wisent 0.7.379__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- wisent/__init__.py +64 -0
- wisent/cli.py +114 -0
- wisent/core/__init__.py +40 -0
- wisent/core/activations/__init__.py +26 -0
- wisent/core/activations/activations.py +97 -0
- wisent/core/activations/activations_collector.py +506 -0
- wisent/core/activations/core/__init__.py +0 -0
- wisent/core/activations/core/atoms.py +219 -0
- wisent/core/activations/prompt_construction_strategy.py +47 -0
- wisent/core/adapters/__init__.py +22 -0
- wisent/core/adapters/audio.py +616 -0
- wisent/core/adapters/base.py +420 -0
- wisent/core/adapters/multimodal.py +738 -0
- wisent/core/adapters/robotics.py +643 -0
- wisent/core/adapters/text.py +441 -0
- wisent/core/adapters/video.py +555 -0
- wisent/core/agent/__init__.py +1 -0
- wisent/core/agent/budget.py +644 -0
- wisent/core/agent/device_benchmarks.py +691 -0
- wisent/core/agent/diagnose/__init__.py +1 -0
- wisent/core/agent/diagnose/agent_classifier_decision.py +641 -0
- wisent/core/agent/diagnose/classifier_marketplace.py +554 -0
- wisent/core/agent/diagnose/create_classifier.py +1155 -0
- wisent/core/agent/diagnose/response_diagnostics.py +273 -0
- wisent/core/agent/diagnose/select_classifiers.py +507 -0
- wisent/core/agent/diagnose/synthetic_classifier_option.py +755 -0
- wisent/core/agent/diagnose/tasks/__init__.py +33 -0
- wisent/core/agent/diagnose/tasks/task_manager.py +1453 -0
- wisent/core/agent/diagnose/tasks/task_relevance.py +94 -0
- wisent/core/agent/diagnose/tasks/task_selector.py +151 -0
- wisent/core/agent/diagnose.py +249 -0
- wisent/core/agent/steer.py +215 -0
- wisent/core/agent/timeout.py +134 -0
- wisent/core/autonomous_agent.py +1158 -0
- wisent/core/benchmark_extractors.py +372 -0
- wisent/core/benchmark_registry.py +151 -0
- wisent/core/bigcode_extractors.py +26 -0
- wisent/core/bigcode_integration.py +886 -0
- wisent/core/branding.py +108 -0
- wisent/core/classifier/__init__.py +1 -0
- wisent/core/classifier/models/__init__.py +1 -0
- wisent/core/classifiers/__init__.py +1 -0
- wisent/core/classifiers/classifiers/__init__.py +0 -0
- wisent/core/classifiers/classifiers/core/__init__.py +0 -0
- wisent/core/classifiers/classifiers/core/atoms.py +748 -0
- wisent/core/classifiers/classifiers/models/__init__.py +0 -0
- wisent/core/classifiers/classifiers/models/logistic.py +29 -0
- wisent/core/classifiers/classifiers/models/mlp.py +47 -0
- wisent/core/classifiers/classifiers/rotator.py +137 -0
- wisent/core/classifiers/core/__init__.py +1 -0
- wisent/core/classifiers/models/__init__.py +1 -0
- wisent/core/classifiers/pipeline_steps/__init__.py +1 -0
- wisent/core/cli/__init__.py +26 -0
- wisent/core/cli/agent/__init__.py +15 -0
- wisent/core/cli/agent/apply_steering.py +192 -0
- wisent/core/cli/agent/evaluate_response.py +128 -0
- wisent/core/cli/agent/generate_synthetic_pairs.py +123 -0
- wisent/core/cli/agent/main.py +139 -0
- wisent/core/cli/agent/train_classifier.py +173 -0
- wisent/core/cli/check_linearity.py +126 -0
- wisent/core/cli/create_steering_vector.py +304 -0
- wisent/core/cli/diagnose_pairs.py +153 -0
- wisent/core/cli/diagnose_vectors.py +404 -0
- wisent/core/cli/estimate_unified_goodness_time.py +428 -0
- wisent/core/cli/evaluate_refusal.py +241 -0
- wisent/core/cli/evaluate_responses.py +926 -0
- wisent/core/cli/generate_humanization_pairs.py +128 -0
- wisent/core/cli/generate_pairs.py +175 -0
- wisent/core/cli/generate_pairs_from_task.py +108 -0
- wisent/core/cli/generate_responses.py +160 -0
- wisent/core/cli/generate_vector_from_synthetic.py +217 -0
- wisent/core/cli/generate_vector_from_task.py +248 -0
- wisent/core/cli/get_activations.py +192 -0
- wisent/core/cli/inference_config.py +84 -0
- wisent/core/cli/inference_config_cli.py +54 -0
- wisent/core/cli/modify_weights.py +660 -0
- wisent/core/cli/multi_steer.py +112 -0
- wisent/core/cli/optimization_cache.py +298 -0
- wisent/core/cli/optimize.py +621 -0
- wisent/core/cli/optimize_classification.py +473 -0
- wisent/core/cli/optimize_sample_size.py +390 -0
- wisent/core/cli/optimize_steering.py +3421 -0
- wisent/core/cli/optimize_weights.py +1287 -0
- wisent/core/cli/steering_method_trainer.py +641 -0
- wisent/core/cli/steering_search_space.py +508 -0
- wisent/core/cli/tasks.py +940 -0
- wisent/core/cli/train_unified_goodness.py +681 -0
- wisent/core/cli_logger.py +22 -0
- wisent/core/config_manager.py +1731 -0
- wisent/core/contrastive_pairs/__init__.py +15 -0
- wisent/core/contrastive_pairs/core/__init__.py +0 -0
- wisent/core/contrastive_pairs/core/atoms.py +45 -0
- wisent/core/contrastive_pairs/core/buliders.py +59 -0
- wisent/core/contrastive_pairs/core/pair.py +183 -0
- wisent/core/contrastive_pairs/core/response.py +153 -0
- wisent/core/contrastive_pairs/core/serialization.py +306 -0
- wisent/core/contrastive_pairs/core/set.py +192 -0
- wisent/core/contrastive_pairs/diagnostics/__init__.py +79 -0
- wisent/core/contrastive_pairs/diagnostics/activations.py +53 -0
- wisent/core/contrastive_pairs/diagnostics/base.py +73 -0
- wisent/core/contrastive_pairs/diagnostics/control_vectors.py +1655 -0
- wisent/core/contrastive_pairs/diagnostics/coverage.py +79 -0
- wisent/core/contrastive_pairs/diagnostics/divergence.py +98 -0
- wisent/core/contrastive_pairs/diagnostics/duplicates.py +118 -0
- wisent/core/contrastive_pairs/diagnostics/linearity.py +325 -0
- wisent/core/contrastive_pairs/diagnostics/vector_quality.py +620 -0
- wisent/core/contrastive_pairs/huggingface_pairs/__init__.py +1 -0
- wisent/core/contrastive_pairs/huggingface_pairs/atoms.py +255 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_extractor_manifest.py +470 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_extractor_registry.py +136 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/__init__.py +44 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentbench.py +225 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentharm.py +267 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentic_search.py +444 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aider_polyglot.py +225 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime.py +118 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime2024.py +74 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime2025.py +73 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/alpaca_eval.py +153 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/apps.py +182 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/arena_hard.py +179 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/atis.py +89 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/babilong.py +96 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bangla_mmlu.py +108 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/basqueglue.py +217 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bec2016eu.py +99 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bfcl.py +283 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bhtc_v2.py +87 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/browsecomp.py +245 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/chain_of_thought.py +89 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/chinese_simpleqa.py +209 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/cluewsc.py +177 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/cnn_dailymail.py +92 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codeforces.py +378 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue.py +109 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text.py +15 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_go.py +64 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_java.py +65 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_javascript.py +65 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_php.py +65 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_python.py +65 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_ruby.py +65 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coding_benchmarks.py +844 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coedit_gec.py +79 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/conala.py +133 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/concode.py +111 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/dbpedia_14.py +91 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/doc_vqa.py +102 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/donotanswer.py +236 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ds1000.py +129 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ds_1000.py +155 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/epec_koref_bin.py +85 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ethos_binary.py +82 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/evalita_mp.py +165 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/evalita_sp_sum_task_fp_small_p1.py +89 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/facts_grounding.py +181 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/faithbench.py +295 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/financial_tweets.py +100 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flames.py +270 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flan_held_in.py +98 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flores.py +572 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/frames.py +143 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/freebase.py +99 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/get_negative_example_livecodebench.py +146 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/get_positive_example_livecodebench.py +140 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/gpt3_translation_benchmarks.py +98 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hallucinations_leaderboard.py +389 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/halueval.py +246 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/harmbench.py +250 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/healthbench.py +181 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hle.py +106 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hmmt.py +117 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/humaneval.py +119 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/humanevalpack.py +102 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/instruct_humaneval.py +180 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/instructhumaneval.py +129 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/iwslt2017_ar_en.py +98 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/iwslt2017_en_ar.py +98 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/jailbreakbench.py +258 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/law_stack_exchange.py +101 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ledgar.py +118 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench.py +61 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench_contrastive_pair_generator.py +491 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench_v6.py +263 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livemathbench.py +230 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/llama.py +96 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/longform_writing.py +285 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/m_mmlu.py +96 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math.py +186 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math500.py +146 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mbpp.py +142 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/meddialog.py +79 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medical_abstracts.py +101 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medium_priority_benchmarks.py +787 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mercury.py +111 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mmlu_redux.py +194 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mmlusr.py +108 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multimedqa.py +99 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multipl_e.py +109 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple.py +96 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_choice.py +87 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_cpp.py +128 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_go.py +128 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_java.py +128 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_js.py +128 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_py.py +15 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_rs.py +128 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/non_greedy_robustness_agieval_aqua_rat.py +92 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/olympiadbench.py +287 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/openllm.py +99 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/option_order_robustness_agieval_aqua_rat.py +92 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/or_bench.py +300 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/penn_treebank.py +80 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/planbench.py +317 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/polymath.py +467 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/prompt_robustness_agieval_aqua_rat.py +92 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/pythia.py +99 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/recode.py +131 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/refusalbench.py +280 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/scicode.py +275 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/self_consistency.py +90 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/simpleqa.py +145 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/sorry_bench.py +211 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/stsb.py +79 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_glue_lm_eval_v1.py +99 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_glue_lm_eval_v1_seq2seq.py +98 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_glue_t5_prompt.py +123 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_gpqa.py +106 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/swe_bench.py +428 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/swe_bench_verified.py +158 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/sycophancy_eval.py +205 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/t0_eval.py +79 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tag.py +98 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tau_bench.py +305 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tmlu.py +109 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolbench.py +360 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolemu.py +386 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/travelplanner.py +286 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/truthfulqa_generation.py +128 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/unfair_tos.py +83 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/vaxx_stance.py +86 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wiceu.py +85 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wikitext103.py +97 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wildguard.py +280 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt14_en_fr.py +97 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt14_fr_en.py +97 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt16_de_en.py +90 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt16_en_de.py +90 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt16_en_ro.py +90 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt16_ro_en.py +90 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt_ro_en_t5_prompt.py +90 -0
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/xsum.py +81 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/__init__.py +0 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/atoms.py +265 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/__init__.py +472 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/aclue.py +24 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/acp.py +33 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/acpbench.py +39 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/advanced_ai_risk.py +59 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/aexams.py +14 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrimgsm.py +10 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrimmlu.py +10 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrixnli.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench.py +14 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_adr.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_afriqa.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_afrisenti.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_belebele.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_flores.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_injongointent.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_mafand.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_masakhaner.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_masakhanews.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_masakhapos.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_naijarc.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_nollysenti.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_ntrex.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_openai_mmlu.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_salt.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_sib.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_uhura_arc_easy.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_xlsum.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/agieval.py +33 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/anli.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arab_culture.py +24 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabic_leaderboard_acva.py +67 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabic_leaderboard_acva_light.py +67 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabic_leaderboard_complete.py +24 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabic_leaderboard_light.py +81 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabicmmlu.py +59 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/aradice.py +36 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arc.py +61 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arithmetic.py +19 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/basque_bench.py +37 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/bbh.py +121 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/bbq.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/belebele.py +293 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/bertaqa.py +25 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/bigbench.py +300 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/blimp.py +76 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/careqa.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/catalan_bench.py +43 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/ceval_valid.py +61 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/cmmlu.py +76 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/code_x_glue.py +16 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/copal_id.py +11 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/crows_pairs.py +31 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/csatqa.py +15 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/darija.py +29 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/darijammlu.py +57 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/egymmlu.py +62 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/eus.py +76 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/evalita_mp.py +93 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/fld.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/flores.py +466 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/freebase.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/french_bench.py +23 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/galician_bench.py +41 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/glianorex.py +11 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/global_mmlu.py +115 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/gpqa.py +27 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/gsm8k.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/gsm8k_platinum.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/haerae.py +14 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/headqa.py +11 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/hellaswag.py +39 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/hendrycks_ethics.py +14 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/hendrycks_math.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/hrm8k.py +20 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/inverse.py +22 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/japanese_leaderboard.py +20 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/jsonschema_bench.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/kbl.py +85 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/kmmlu.py +281 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/kobest.py +14 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/kormedmcqa.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/lambada.py +28 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/leaderboard.py +52 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/libra.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/lingoly.py +11 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/longbench.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/m.py +43 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mastermind.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mathqa.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/med.py +24 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/meddialog.py +12 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/medqa.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mela.py +18 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/metabench.py +36 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mgsm.py +44 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/minerva_math.py +16 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mlqa.py +58 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlu.py +70 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlu_pro.py +23 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlu_pro_plus.py +23 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlu_prox.py +191 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlusr.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmmu.py +46 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/model_written_evals.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/multiblimp.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/non.py +23 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/noreval.py +143 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/noridiom.py +20 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/nortruthfulqa.py +32 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/nrk.py +20 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi_arc_multilingual.py +10 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi_hellaswag_multilingual.py +24 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi_mmlu_multilingual.py +24 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi_truthfulqa_multilingual.py +34 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/paloma.py +25 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/pawsx.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/persona.py +144 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/pile.py +31 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/polemo2.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/portuguese_bench.py +31 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/prompt.py +23 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/qa4mre.py +12 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/qasper.py +11 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/ru.py +19 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/ruler.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/score.py +20 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/scrolls.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/self_consistency.py +11 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/spanish_bench.py +38 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/storycloze.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/super_glue_t5_prompt.py +17 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/tinyBenchmarks.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/tmlu.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/tmmluplus.py +80 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/translation.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/truthfulqa.py +76 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/truthfulqa_multi.py +24 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/turkishmmlu.py +30 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/unitxt.py +23 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/unscramble.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/winogender.py +16 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/wmdp.py +12 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/wmt14.py +16 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/wmt16.py +22 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/wsc273.py +9 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xcopa.py +21 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xnli.py +28 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xnli_eu.py +12 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xquad.py +22 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xstorycloze.py +22 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xwinograd.py +15 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_manifest.py +478 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_registry.py +140 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/__init__.py +125 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aclue.py +171 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench.py +207 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench_hard.py +185 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/advanced.py +130 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aexams.py +184 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrimgsm.py +98 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrimmlu.py +113 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrixnli.py +129 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrobench_cot.py +88 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrobench_mc.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ag.py +134 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/agieval.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ai2_arc.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anagrams1.py +81 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anagrams2.py +81 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anli.py +140 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabculture.py +180 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic.py +98 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_exams.py +104 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_complete.py +168 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_light.py +168 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabicmmlu.py +167 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aradice.py +268 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc.py +133 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_challenge.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_easy.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_gen.py +101 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_mc.py +106 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/argument.py +134 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arithmetic.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/asdiv.py +122 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/assin.py +103 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/babi.py +113 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench_gen.py +168 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench_mc.py +139 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bbh.py +133 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bbq.py +169 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/belebele.py +181 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/benchmarks.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bertaqa.py +165 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhs.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhtc.py +143 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bigbench.py +170 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp.py +171 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp_nl.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/boolq.py +117 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/boolq_seq2seq.py +117 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/c4.py +150 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cabbq.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cabreu.py +127 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/careqa.py +169 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench_gen.py +119 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench_mc.py +113 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalanqa.py +171 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catcola.py +139 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cb.py +117 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval.py +223 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval_valid.py +163 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chain.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chartqa.py +238 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/claim.py +151 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/click.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cmmlu.py +166 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cnn.py +144 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cocoteros.py +148 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code2text.py +161 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code_x_glue.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/codexglue.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coedit.py +149 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cola.py +83 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense_qa.py +127 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copa.py +124 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copal_id.py +169 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coqa.py +162 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coqcat.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/crows_pairs.py +158 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/csatqa.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cycle.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cycle_letters.py +81 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darija_bench.py +221 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijahellaswag.py +174 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijammlu.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/dbpedia.py +157 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/discrim_eval.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/doc.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/drop.py +129 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/egyhellaswag.py +125 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/egymmlu.py +180 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/epec.py +142 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench.py +194 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_ca.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_es.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/esbbq.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/escola.py +85 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ethics.py +135 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ethos.py +99 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_exams.py +225 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_proficiency.py +159 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_reading.py +159 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_trivia.py +159 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/evalita_llm.py +166 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/evalita_sp.py +109 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/fda.py +105 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/financial.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/flan.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/fld.py +143 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench.py +202 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench_mc.py +98 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench_perplexity.py +86 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galcola.py +109 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench_gen.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench_mc.py +112 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gaokao.py +141 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/glianorex.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_mmlu.py +171 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_piqa.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/glue.py +109 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpqa.py +161 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpt3.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/groundcocoa.py +184 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gsm.py +108 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gsm8k.py +134 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/haerae.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/headqa.py +112 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hellaswag.py +125 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_ethics.py +225 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_math.py +191 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/histoires_morales.py +179 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hle.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hrm8k.py +203 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/humaneval.py +124 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/humaneval_infilling.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/icelandic_winogrande.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ifeval.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse_scaling.py +192 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/iwslt2017.py +117 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ja.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard_gen.py +224 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard_mc.py +120 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/jsonschema_bench.py +123 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kbl.py +140 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu.py +168 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu_cot.py +88 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu_mc.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kobest.py +165 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kormedmcqa.py +160 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada.py +147 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_cloze.py +185 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_multilingual.py +185 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_multilingual_stablelm.py +141 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/law.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/leaderboard.py +194 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/libra.py +165 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lingoly.py +203 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/livemathbench.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/llama3.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lm_syneval.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logieval.py +82 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa.py +115 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa2.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbench.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbenchv2.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mastermind.py +203 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mathqa.py +137 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mbpp.py +123 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mc-taco.py +115 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/med_concepts_qa.py +224 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meddialog.py +180 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medical.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mediqa_qa2019.py +123 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medmcqa.py +169 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medqa.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medtext.py +108 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mela.py +96 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meqsum.py +115 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/metabench.py +154 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mgsm.py +122 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mimic_repsum.py +140 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/minerva_math.py +172 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mlqa.py +143 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu.py +144 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_cot.py +88 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_mc.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_pro.py +145 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlusr.py +189 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmmu.py +150 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mnli.py +113 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/model_written_evals.py +115 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/moral_stories.py +151 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mrpc.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mts_dialog.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mts_dialog_perplexity.py +97 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multiblimp.py +134 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multilingual.py +106 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multirc.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mutual.py +113 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/non.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval.py +173 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_exact.py +157 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_gen.py +277 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_gen_exact.py +165 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc.py +228 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc_log_likelihoods.py +223 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noticia.py +105 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/nq_open.py +135 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi.py +27 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_arc_multilingual.py +167 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_hellaswag_multilingual.py +174 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_mmlu_multilingual.py +162 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_truthfulqa_multilingual.py +209 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/olaph.py +186 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/olaph_perplexity.py +97 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/openbookqa.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/option.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paloma.py +205 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafraseja.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafrases.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws_x.py +154 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pawsx.py +115 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/persona.py +246 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/phrases.py +144 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/phrases_ca_va.py +82 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile.py +161 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile_10k.py +140 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/piqa.py +116 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/polemo2.py +135 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/polymath.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench_gen.py +121 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench_mc.py +103 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prompt.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prost.py +115 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pubmedqa.py +112 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qa4mre.py +119 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper_bool.py +112 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnli.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnlieu.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qqp.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/quac.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/race.py +124 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/random.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/realtoxicityprompts.py +124 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/record.py +125 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/reversed.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/rte.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ruler.py +170 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sciq.py +113 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/score.py +177 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls.py +161 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls_mc.py +157 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/self.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue.py +131 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue_rte.py +119 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/simple_cooccurrence_bias.py +121 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/siqa.py +209 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/social_iqa.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench_gen.py +117 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench_mc.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/squad2.py +129 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/squad_completion.py +121 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sst2.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/storycloze.py +250 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/summarization.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super.py +107 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super_glue.py +154 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/superglue.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/supergpqa.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swag.py +115 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swde.py +179 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sycophancy.py +117 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/t0.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/teca.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyarc.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinybenchmarks.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinygsm8k.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyhellaswag.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinymmlu.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinytruthfulqa.py +113 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinywinogrande.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tmmluplus.py +181 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/toxigen.py +91 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/translation.py +149 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/triviaqa.py +130 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa.py +112 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc1.py +120 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc2.py +140 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_multi.py +142 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turblimp_core.py +152 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu.py +161 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu_cot.py +104 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu_mc.py +102 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/twenty_newsgroups.py +111 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/unitxt.py +131 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/unscramble.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/vaxx.py +95 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/webqs.py +130 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wic.py +122 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wikitext.py +146 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/winogender.py +139 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/winogrande.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmdp.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmt14.py +110 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmt16.py +118 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wnli.py +114 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc.py +117 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc273.py +180 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xcopa.py +197 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xlsum.py +147 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xnli.py +131 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xquad.py +203 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xstorycloze.py +129 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xwinograd.py +124 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/yahoo.py +108 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/zhoblimp.py +155 -0
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_pairs_generation.py +56 -0
- wisent/core/data_loaders/__init__.py +235 -0
- wisent/core/data_loaders/core/__init__.py +0 -0
- wisent/core/data_loaders/core/atoms.py +99 -0
- wisent/core/data_loaders/loaders/__init__.py +0 -0
- wisent/core/data_loaders/loaders/custom.py +120 -0
- wisent/core/data_loaders/loaders/huggingface_loader.py +153 -0
- wisent/core/data_loaders/loaders/lm_loader.py +494 -0
- wisent/core/data_loaders/loaders/lm_loader_special_cases.py +496 -0
- wisent/core/data_loaders/loaders/task_interface_loader.py +300 -0
- wisent/core/data_loaders/rotator.py +118 -0
- wisent/core/detection_handling.py +259 -0
- wisent/core/diversity_processors.py +193 -0
- wisent/core/download_full_benchmarks.py +1512 -0
- wisent/core/errors/__init__.py +203 -0
- wisent/core/errors/error_codes.py +763 -0
- wisent/core/errors/error_handler.py +134 -0
- wisent/core/evaluators/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/__init__.py +42 -0
- wisent/core/evaluators/benchmark_specific/aime_evaluator.py +90 -0
- wisent/core/evaluators/benchmark_specific/coding/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/coding/metrics/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/coding/metrics/core/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/coding/metrics/core/atoms.py +36 -0
- wisent/core/evaluators/benchmark_specific/coding/metrics/evaluator.py +363 -0
- wisent/core/evaluators/benchmark_specific/coding/metrics/passk.py +67 -0
- wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/core/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/core/atoms.py +27 -0
- wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/cpp_sanitizer.py +62 -0
- wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/java_sanitizer.py +78 -0
- wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/python_sanitizer.py +94 -0
- wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/utils.py +126 -0
- wisent/core/evaluators/benchmark_specific/coding/providers/__init__.py +18 -0
- wisent/core/evaluators/benchmark_specific/coding/providers/core/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/coding/providers/core/atoms.py +31 -0
- wisent/core/evaluators/benchmark_specific/coding/providers/livecodebench/__init__.py +3 -0
- wisent/core/evaluators/benchmark_specific/coding/providers/livecodebench/provider.py +305 -0
- wisent/core/evaluators/benchmark_specific/coding/safe_docker/Dockerfile +31 -0
- wisent/core/evaluators/benchmark_specific/coding/safe_docker/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/coding/safe_docker/core/__init__.py +0 -0
- wisent/core/evaluators/benchmark_specific/coding/safe_docker/core/atoms.py +105 -0
- wisent/core/evaluators/benchmark_specific/coding/safe_docker/core/runtime.py +143 -0
- wisent/core/evaluators/benchmark_specific/coding/safe_docker/entrypoint.py +121 -0
- wisent/core/evaluators/benchmark_specific/coding/safe_docker/recipes.py +60 -0
- wisent/core/evaluators/benchmark_specific/coding/solution_generator.py +258 -0
- wisent/core/evaluators/benchmark_specific/conala_evaluator.py +332 -0
- wisent/core/evaluators/benchmark_specific/exact_match_evaluator.py +81 -0
- wisent/core/evaluators/benchmark_specific/f1_evaluator.py +173 -0
- wisent/core/evaluators/benchmark_specific/generation_evaluator.py +488 -0
- wisent/core/evaluators/benchmark_specific/livemathbench_evaluator.py +393 -0
- wisent/core/evaluators/benchmark_specific/log_likelihoods_evaluator.py +202 -0
- wisent/core/evaluators/benchmark_specific/math_evaluator.py +119 -0
- wisent/core/evaluators/benchmark_specific/math_parsing/__init__.py +1 -0
- wisent/core/evaluators/benchmark_specific/math_parsing/core.py +1640 -0
- wisent/core/evaluators/benchmark_specific/math_parsing/extract_boxed.py +48 -0
- wisent/core/evaluators/benchmark_specific/math_parsing/is_equiv.py +159 -0
- wisent/core/evaluators/benchmark_specific/math_parsing/scripts.py +919 -0
- wisent/core/evaluators/benchmark_specific/perplexity_evaluator.py +175 -0
- wisent/core/evaluators/benchmark_specific/polymath_evaluator.py +114 -0
- wisent/core/evaluators/core/__init__.py +5 -0
- wisent/core/evaluators/core/atoms.py +166 -0
- wisent/core/evaluators/custom/__init__.py +20 -0
- wisent/core/evaluators/custom/custom_evaluator.py +382 -0
- wisent/core/evaluators/custom/examples/__init__.py +37 -0
- wisent/core/evaluators/custom/examples/desklib_detector.py +166 -0
- wisent/core/evaluators/custom/examples/gptzero.py +185 -0
- wisent/core/evaluators/custom/examples/humanization.py +79 -0
- wisent/core/evaluators/custom/examples/humanization_coherent.py +127 -0
- wisent/core/evaluators/custom/examples/roberta_detector.py +173 -0
- wisent/core/evaluators/oracles/__init__.py +0 -0
- wisent/core/evaluators/oracles/interactive.py +73 -0
- wisent/core/evaluators/oracles/nlp_evaluator.py +440 -0
- wisent/core/evaluators/oracles/truthfulqa_gen_evaluator.py +168 -0
- wisent/core/evaluators/oracles/user_specified.py +67 -0
- wisent/core/evaluators/personalization/__init__.py +12 -0
- wisent/core/evaluators/personalization/alignment.py +166 -0
- wisent/core/evaluators/personalization/coherence.py +325 -0
- wisent/core/evaluators/personalization/difference.py +73 -0
- wisent/core/evaluators/rotator.py +217 -0
- wisent/core/evaluators/steering_evaluators.py +386 -0
- wisent/core/evaluators/synthetic_evaluator.py +377 -0
- wisent/core/hyperparameter_optimizer.py +547 -0
- wisent/core/layer.py +17 -0
- wisent/core/lm_eval_harness_ground_truth.py +1431 -0
- wisent/core/main.py +101 -0
- wisent/core/managed_cached_benchmarks.py +609 -0
- wisent/core/mixed_benchmark_sampler.py +366 -0
- wisent/core/modalities/__init__.py +545 -0
- wisent/core/model_persistence.py +302 -0
- wisent/core/models/__init__.py +23 -0
- wisent/core/models/core/__init__.py +0 -0
- wisent/core/models/core/atoms.py +465 -0
- wisent/core/models/inference_config.py +127 -0
- wisent/core/models/wisent_model.py +893 -0
- wisent/core/multi_steering.py +397 -0
- wisent/core/opti/__init__.py +0 -0
- wisent/core/opti/core/__init__.py +0 -0
- wisent/core/opti/core/atoms.py +177 -0
- wisent/core/opti/methods/__init__.py +10 -0
- wisent/core/opti/methods/opti_classificator.py +172 -0
- wisent/core/opti/methods/opti_steering.py +139 -0
- wisent/core/opti/methods/opti_weights.py +523 -0
- wisent/core/optuna/__init__.py +54 -0
- wisent/core/optuna/classifier/__init__.py +25 -0
- wisent/core/optuna/classifier/activation_generator.py +351 -0
- wisent/core/optuna/classifier/classifier_cache.py +509 -0
- wisent/core/optuna/classifier/optuna_classifier_optimizer.py +685 -0
- wisent/core/optuna/steering/__init__.py +20 -0
- wisent/core/optuna/steering/bigcode_evaluator_wrapper.py +200 -0
- wisent/core/optuna/steering/data_utils.py +342 -0
- wisent/core/optuna/steering/metrics.py +412 -0
- wisent/core/optuna/steering/steering_optimization.py +1096 -0
- wisent/core/parser.py +1662 -0
- wisent/core/parser_arguments/__init__.py +10 -0
- wisent/core/parser_arguments/agent_parser.py +122 -0
- wisent/core/parser_arguments/check_linearity_parser.py +82 -0
- wisent/core/parser_arguments/configure_model_parser.py +7 -0
- wisent/core/parser_arguments/create_steering_vector_parser.py +67 -0
- wisent/core/parser_arguments/diagnose_pairs_parser.py +25 -0
- wisent/core/parser_arguments/diagnose_vectors_parser.py +72 -0
- wisent/core/parser_arguments/evaluate_parser.py +40 -0
- wisent/core/parser_arguments/evaluate_refusal_parser.py +32 -0
- wisent/core/parser_arguments/evaluate_responses_parser.py +12 -0
- wisent/core/parser_arguments/full_optimize_parser.py +194 -0
- wisent/core/parser_arguments/generate_pairs_from_task_parser.py +33 -0
- wisent/core/parser_arguments/generate_pairs_parser.py +43 -0
- wisent/core/parser_arguments/generate_responses_parser.py +16 -0
- wisent/core/parser_arguments/generate_vector_from_synthetic_parser.py +148 -0
- wisent/core/parser_arguments/generate_vector_from_task_parser.py +149 -0
- wisent/core/parser_arguments/generate_vector_parser.py +89 -0
- wisent/core/parser_arguments/get_activations_parser.py +90 -0
- wisent/core/parser_arguments/inference_config_parser.py +65 -0
- wisent/core/parser_arguments/main_parser.py +220 -0
- wisent/core/parser_arguments/model_config_parser.py +59 -0
- wisent/core/parser_arguments/modify_weights_parser.py +309 -0
- wisent/core/parser_arguments/monitor_parser.py +17 -0
- wisent/core/parser_arguments/multi_steer_parser.py +48 -0
- wisent/core/parser_arguments/nonsense_parser.py +26 -0
- wisent/core/parser_arguments/optimization_cache_parser.py +64 -0
- wisent/core/parser_arguments/optimize_classification_parser.py +108 -0
- wisent/core/parser_arguments/optimize_parser.py +142 -0
- wisent/core/parser_arguments/optimize_sample_size_parser.py +58 -0
- wisent/core/parser_arguments/optimize_steering_parser.py +617 -0
- wisent/core/parser_arguments/optimize_weights_parser.py +403 -0
- wisent/core/parser_arguments/synthetic_parser.py +117 -0
- wisent/core/parser_arguments/tasks_parser.py +591 -0
- wisent/core/parser_arguments/train_unified_goodness_parser.py +172 -0
- wisent/core/parser_arguments/utils.py +107 -0
- wisent/core/prompts/__init__.py +0 -0
- wisent/core/prompts/core/__init__.py +0 -0
- wisent/core/prompts/core/atom.py +57 -0
- wisent/core/prompts/core/prompt_formater.py +148 -0
- wisent/core/prompts/prompt_stratiegies/__init__.py +0 -0
- wisent/core/prompts/prompt_stratiegies/direct_completion.py +26 -0
- wisent/core/prompts/prompt_stratiegies/instruction_following.py +26 -0
- wisent/core/prompts/prompt_stratiegies/multiple_choice.py +31 -0
- wisent/core/prompts/prompt_stratiegies/role_playing.py +33 -0
- wisent/core/representation.py +5 -0
- wisent/core/save_results.py +277 -0
- wisent/core/steering.py +660 -0
- wisent/core/steering_method.py +20 -0
- wisent/core/steering_methods/__init__.py +54 -0
- wisent/core/steering_methods/core/__init__.py +0 -0
- wisent/core/steering_methods/core/atoms.py +154 -0
- wisent/core/steering_methods/methods/__init__.py +0 -0
- wisent/core/steering_methods/methods/caa.py +45 -0
- wisent/core/steering_methods/methods/prism.py +588 -0
- wisent/core/steering_methods/methods/pulse.py +641 -0
- wisent/core/steering_methods/methods/titan.py +1005 -0
- wisent/core/steering_methods/preflight.py +322 -0
- wisent/core/steering_methods/registry.py +649 -0
- wisent/core/steering_methods/rotator.py +121 -0
- wisent/core/steering_optimizer.py +1503 -0
- wisent/core/synthetic/__init__.py +0 -0
- wisent/core/synthetic/cleaners/__init__.py +0 -0
- wisent/core/synthetic/cleaners/core/__init__.py +0 -0
- wisent/core/synthetic/cleaners/core/atoms.py +58 -0
- wisent/core/synthetic/cleaners/deduper_cleaner.py +53 -0
- wisent/core/synthetic/cleaners/methods/__init__.py +0 -0
- wisent/core/synthetic/cleaners/methods/base_dedupers.py +321 -0
- wisent/core/synthetic/cleaners/methods/base_refusalers.py +286 -0
- wisent/core/synthetic/cleaners/methods/core/__init__.py +0 -0
- wisent/core/synthetic/cleaners/methods/core/atoms.py +47 -0
- wisent/core/synthetic/cleaners/pairs_cleaner.py +90 -0
- wisent/core/synthetic/cleaners/refusaler_cleaner.py +133 -0
- wisent/core/synthetic/db_instructions/__init__.py +0 -0
- wisent/core/synthetic/db_instructions/core/__init__.py +0 -0
- wisent/core/synthetic/db_instructions/core/atoms.py +25 -0
- wisent/core/synthetic/db_instructions/mini_dp.py +115 -0
- wisent/core/synthetic/generators/__init__.py +0 -0
- wisent/core/synthetic/generators/core/__init__.py +0 -0
- wisent/core/synthetic/generators/core/atoms.py +73 -0
- wisent/core/synthetic/generators/diversities/__init__.py +0 -0
- wisent/core/synthetic/generators/diversities/core/__init__.py +0 -0
- wisent/core/synthetic/generators/diversities/core/core.py +68 -0
- wisent/core/synthetic/generators/diversities/methods/__init__.py +0 -0
- wisent/core/synthetic/generators/diversities/methods/fast_diversity.py +249 -0
- wisent/core/synthetic/generators/nonsense_generator.py +150 -0
- wisent/core/synthetic/generators/pairs_generator.py +313 -0
- wisent/core/task_interface.py +143 -0
- wisent/core/task_selector.py +232 -0
- wisent/core/tasks/__init__.py +218 -0
- wisent/core/tasks/aime_task.py +142 -0
- wisent/core/tasks/file_task.py +212 -0
- wisent/core/tasks/hle_task.py +180 -0
- wisent/core/tasks/hmmt_task.py +120 -0
- wisent/core/tasks/livecodebench_task.py +94 -0
- wisent/core/tasks/livemathbench_task.py +159 -0
- wisent/core/tasks/lm_eval_task.py +611 -0
- wisent/core/tasks/math500_task.py +84 -0
- wisent/core/tasks/polymath_task.py +147 -0
- wisent/core/tasks/supergpqa_task.py +220 -0
- wisent/core/time_estimator.py +155 -0
- wisent/core/timing_calibration.py +176 -0
- wisent/core/tracking/__init__.py +54 -0
- wisent/core/tracking/latency.py +620 -0
- wisent/core/tracking/memory.py +360 -0
- wisent/core/trainers/__init__.py +0 -0
- wisent/core/trainers/core/__init__.py +11 -0
- wisent/core/trainers/core/atoms.py +45 -0
- wisent/core/trainers/steering_trainer.py +365 -0
- wisent/core/universal_subspace.py +918 -0
- wisent/core/user_model_config.py +158 -0
- wisent/core/utils/__init__.py +64 -0
- wisent/core/utils/base_rotator.py +292 -0
- wisent/core/utils/dataset_splits.py +197 -0
- wisent/core/utils/device.py +279 -0
- wisent/core/weight_modification/__init__.py +134 -0
- wisent/core/weight_modification/additive.py +340 -0
- wisent/core/weight_modification/directional.py +1357 -0
- wisent/core/weight_modification/export.py +359 -0
- wisent/core/weight_modification/multi_direction.py +410 -0
- wisent/core/weight_modification/utils.py +236 -0
- wisent/core/wisent.py +660 -0
- wisent/examples/contrastive_pairs/humanization_human_vs_ai.json +2112 -0
- wisent/examples/scripts/1/test_basqueglue_evaluation.json +51 -0
- wisent/examples/scripts/1/test_basqueglue_pairs.json +14 -0
- wisent/examples/scripts/1/test_bec2016eu_evaluation.json +51 -0
- wisent/examples/scripts/1/test_bec2016eu_pairs.json +14 -0
- wisent/examples/scripts/1/test_belebele_evaluation.json +51 -0
- wisent/examples/scripts/1/test_belebele_pairs.json +14 -0
- wisent/examples/scripts/1/test_benchmarks_evaluation.json +51 -0
- wisent/examples/scripts/1/test_benchmarks_pairs.json +14 -0
- wisent/examples/scripts/1/test_bertaqa_evaluation.json +51 -0
- wisent/examples/scripts/1/test_bertaqa_pairs.json +14 -0
- wisent/examples/scripts/1/test_bhtc_v2_evaluation.json +30 -0
- wisent/examples/scripts/1/test_bhtc_v2_pairs.json +8 -0
- wisent/examples/scripts/1/test_boolq-seq2seq_evaluation.json +30 -0
- wisent/examples/scripts/1/test_boolq-seq2seq_pairs.json +8 -0
- wisent/examples/scripts/1/test_cabreu_evaluation.json +30 -0
- wisent/examples/scripts/1/test_cabreu_pairs.json +8 -0
- wisent/examples/scripts/1/test_careqa_en_evaluation.json +30 -0
- wisent/examples/scripts/1/test_careqa_en_pairs.json +8 -0
- wisent/examples/scripts/1/test_careqa_evaluation.json +30 -0
- wisent/examples/scripts/1/test_careqa_pairs.json +8 -0
- wisent/examples/scripts/1/test_catalanqa_evaluation.json +30 -0
- wisent/examples/scripts/1/test_catalanqa_pairs.json +8 -0
- wisent/examples/scripts/1/test_catcola_evaluation.json +30 -0
- wisent/examples/scripts/1/test_catcola_pairs.json +8 -0
- wisent/examples/scripts/1/test_chartqa_evaluation.json +30 -0
- wisent/examples/scripts/1/test_chartqa_pairs.json +8 -0
- wisent/examples/scripts/1/test_claim_stance_topic_evaluation.json +30 -0
- wisent/examples/scripts/1/test_claim_stance_topic_pairs.json +8 -0
- wisent/examples/scripts/1/test_cnn_dailymail_evaluation.json +30 -0
- wisent/examples/scripts/1/test_cnn_dailymail_pairs.json +8 -0
- wisent/examples/scripts/1/test_cocoteros_es_evaluation.json +30 -0
- wisent/examples/scripts/1/test_cocoteros_es_pairs.json +8 -0
- wisent/examples/scripts/1/test_coedit_gec_evaluation.json +30 -0
- wisent/examples/scripts/1/test_coedit_gec_pairs.json +8 -0
- wisent/examples/scripts/1/test_cola_evaluation.json +30 -0
- wisent/examples/scripts/1/test_cola_pairs.json +8 -0
- wisent/examples/scripts/1/test_coqcat_evaluation.json +30 -0
- wisent/examples/scripts/1/test_coqcat_pairs.json +8 -0
- wisent/examples/scripts/1/test_dbpedia_14_evaluation.json +30 -0
- wisent/examples/scripts/1/test_dbpedia_14_pairs.json +8 -0
- wisent/examples/scripts/1/test_epec_koref_bin_evaluation.json +30 -0
- wisent/examples/scripts/1/test_epec_koref_bin_pairs.json +8 -0
- wisent/examples/scripts/1/test_ethos_binary_evaluation.json +30 -0
- wisent/examples/scripts/1/test_ethos_binary_pairs.json +8 -0
- wisent/examples/scripts/2/test_afrimgsm_direct_amh_evaluation.json +30 -0
- wisent/examples/scripts/2/test_afrimgsm_direct_amh_pairs.json +8 -0
- wisent/examples/scripts/2/test_afrimmlu_direct_amh_evaluation.json +30 -0
- wisent/examples/scripts/2/test_afrimmlu_direct_amh_pairs.json +8 -0
- wisent/examples/scripts/2/test_afrixnli_en_direct_amh_evaluation.json +30 -0
- wisent/examples/scripts/2/test_afrixnli_en_direct_amh_pairs.json +8 -0
- wisent/examples/scripts/2/test_arc_ar_evaluation.json +30 -0
- wisent/examples/scripts/2/test_arc_ar_pairs.json +8 -0
- wisent/examples/scripts/2/test_atis_evaluation.json +30 -0
- wisent/examples/scripts/2/test_atis_pairs.json +8 -0
- wisent/examples/scripts/2/test_babi_evaluation.json +30 -0
- wisent/examples/scripts/2/test_babi_pairs.json +8 -0
- wisent/examples/scripts/2/test_babilong_evaluation.json +30 -0
- wisent/examples/scripts/2/test_babilong_pairs.json +8 -0
- wisent/examples/scripts/2/test_bangla_mmlu_evaluation.json +30 -0
- wisent/examples/scripts/2/test_bangla_mmlu_pairs.json +8 -0
- wisent/examples/scripts/2/test_basque-glue_pairs.json +14 -0
- wisent/examples/scripts/benchmark_tags.json +2140 -0
- wisent/examples/scripts/lm_eval_readme.json +4 -0
- wisent/examples/scripts/results/benchmark_descriptions.json +1244 -0
- wisent/examples/scripts/results/benchmark_evaluation_methods.json +66 -0
- wisent/examples/scripts/results/benchmark_evaluator_mapping.json +2781 -0
- wisent/examples/scripts/results/benchmark_evaluator_mapping_updated.json +30536 -0
- wisent/examples/scripts/results/benchmark_evaluators_clean.json +469 -0
- wisent/examples/scripts/results/benchmark_methods_summary.json +260 -0
- wisent/examples/scripts/results/benchmark_pair_creation_methods.json +66 -0
- wisent/examples/scripts/results/benchmark_pair_totals.json +269 -0
- wisent/examples/scripts/results/benchmark_tags.json +917 -0
- wisent/examples/scripts/results/benchmark_test_summary_nov4.json +71 -0
- wisent/examples/scripts/results/coding_benchmarks_test_code_status.json +150 -0
- wisent/examples/scripts/results/failing_benchmarks.json +946 -0
- wisent/examples/scripts/results/failing_benchmarks_list.json +41 -0
- wisent/examples/scripts/results/failing_benchmarks_test_results.json +945 -0
- wisent/examples/scripts/results/missing_benchmark_tags.json +341 -0
- wisent/examples/scripts/results/test_20_newsgroups_evaluation.json +30 -0
- wisent/examples/scripts/results/test_20_newsgroups_pairs.json +8 -0
- wisent/examples/scripts/results/test_AraDICE_evaluation.json +51 -0
- wisent/examples/scripts/results/test_AraDICE_pairs.json +14 -0
- wisent/examples/scripts/results/test_AraDiCE_boolq_egy/test_AraDiCE_boolq_egy_evaluation.json +30 -0
- wisent/examples/scripts/results/test_AraDiCE_boolq_egy/test_AraDiCE_boolq_egy_pairs.json +8 -0
- wisent/examples/scripts/results/test_ArabCulture_evaluation.json +51 -0
- wisent/examples/scripts/results/test_ArabCulture_pairs.json +14 -0
- wisent/examples/scripts/results/test_Tag_evaluation.json +30 -0
- wisent/examples/scripts/results/test_Tag_pairs.json +8 -0
- wisent/examples/scripts/results/test_aclue_evaluation.json +51 -0
- wisent/examples/scripts/results/test_aclue_pairs.json +14 -0
- wisent/examples/scripts/results/test_acp_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_acp_bench_hard_evaluation.json +51 -0
- wisent/examples/scripts/results/test_acp_bench_hard_pairs.json +14 -0
- wisent/examples/scripts/results/test_acp_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_advanced_ai_risk_evaluation.json +51 -0
- wisent/examples/scripts/results/test_advanced_ai_risk_pairs.json +14 -0
- wisent/examples/scripts/results/test_aexams_evaluation.json +51 -0
- wisent/examples/scripts/results/test_aexams_pairs.json +14 -0
- wisent/examples/scripts/results/test_afrimgsm_direct_amh_evaluation.json +30 -0
- wisent/examples/scripts/results/test_afrimgsm_direct_amh_pairs.json +8 -0
- wisent/examples/scripts/results/test_afrimmlu_direct_amh_evaluation.json +30 -0
- wisent/examples/scripts/results/test_afrimmlu_direct_amh_pairs.json +8 -0
- wisent/examples/scripts/results/test_afrixnli_en_direct_amh_evaluation.json +30 -0
- wisent/examples/scripts/results/test_afrixnli_en_direct_amh_pairs.json +8 -0
- wisent/examples/scripts/results/test_ag_news_evaluation.json +30 -0
- wisent/examples/scripts/results/test_ag_news_pairs.json +8 -0
- wisent/examples/scripts/results/test_agieval_evaluation.json +51 -0
- wisent/examples/scripts/results/test_agieval_pairs.json +14 -0
- wisent/examples/scripts/results/test_aime2024_evaluation.json +30 -0
- wisent/examples/scripts/results/test_aime2024_pairs.json +8 -0
- wisent/examples/scripts/results/test_aime2025_evaluation.json +30 -0
- wisent/examples/scripts/results/test_aime2025_pairs.json +8 -0
- wisent/examples/scripts/results/test_aime_evaluation.json +30 -0
- wisent/examples/scripts/results/test_aime_pairs.json +8 -0
- wisent/examples/scripts/results/test_anagrams1_evaluation.json +30 -0
- wisent/examples/scripts/results/test_anagrams1_pairs.json +8 -0
- wisent/examples/scripts/results/test_anagrams2_evaluation.json +30 -0
- wisent/examples/scripts/results/test_anagrams2_pairs.json +8 -0
- wisent/examples/scripts/results/test_anli_evaluation.json +30 -0
- wisent/examples/scripts/results/test_anli_pairs.json +8 -0
- wisent/examples/scripts/results/test_apps_evaluation.json +30 -0
- wisent/examples/scripts/results/test_apps_pairs.json +8 -0
- wisent/examples/scripts/results/test_arabic_exams_evaluation.json +30 -0
- wisent/examples/scripts/results/test_arabic_exams_pairs.json +8 -0
- wisent/examples/scripts/results/test_arabic_leaderboard_complete_evaluation.json +51 -0
- wisent/examples/scripts/results/test_arabic_leaderboard_complete_pairs.json +14 -0
- wisent/examples/scripts/results/test_arabic_leaderboard_light_evaluation.json +51 -0
- wisent/examples/scripts/results/test_arabic_leaderboard_light_pairs.json +14 -0
- wisent/examples/scripts/results/test_arabicmmlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_arabicmmlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_aradice/test_aradice_evaluation.json +51 -0
- wisent/examples/scripts/results/test_aradice/test_aradice_pairs.json +14 -0
- wisent/examples/scripts/results/test_aradice3/test_aradice_evaluation.json +51 -0
- wisent/examples/scripts/results/test_aradice3/test_aradice_pairs.json +14 -0
- wisent/examples/scripts/results/test_arc_ar_evaluation.json +30 -0
- wisent/examples/scripts/results/test_arc_ar_pairs.json +8 -0
- wisent/examples/scripts/results/test_arc_challenge_evaluation.json +30 -0
- wisent/examples/scripts/results/test_arc_challenge_pairs.json +8 -0
- wisent/examples/scripts/results/test_arc_easy_evaluation.json +30 -0
- wisent/examples/scripts/results/test_arc_easy_pairs.json +8 -0
- wisent/examples/scripts/results/test_argument_topic_evaluation.json +30 -0
- wisent/examples/scripts/results/test_argument_topic_pairs.json +8 -0
- wisent/examples/scripts/results/test_arithmetic_evaluation.json +51 -0
- wisent/examples/scripts/results/test_arithmetic_pairs.json +14 -0
- wisent/examples/scripts/results/test_asdiv_evaluation.json +30 -0
- wisent/examples/scripts/results/test_asdiv_pairs.json +8 -0
- wisent/examples/scripts/results/test_assin_entailment_evaluation.json +30 -0
- wisent/examples/scripts/results/test_assin_entailment_pairs.json +8 -0
- wisent/examples/scripts/results/test_atis_evaluation.json +30 -0
- wisent/examples/scripts/results/test_atis_pairs.json +8 -0
- wisent/examples/scripts/results/test_babi_evaluation.json +30 -0
- wisent/examples/scripts/results/test_babi_pairs.json +8 -0
- wisent/examples/scripts/results/test_babilong_evaluation.json +30 -0
- wisent/examples/scripts/results/test_babilong_pairs.json +8 -0
- wisent/examples/scripts/results/test_bangla_mmlu_evaluation.json +30 -0
- wisent/examples/scripts/results/test_bangla_mmlu_pairs.json +8 -0
- wisent/examples/scripts/results/test_banking77_evaluation.json +30 -0
- wisent/examples/scripts/results/test_banking77_pairs.json +8 -0
- wisent/examples/scripts/results/test_basque/test_basque-glue_pairs.json +14 -0
- wisent/examples/scripts/results/test_basque-glue_evaluation.json +51 -0
- wisent/examples/scripts/results/test_basque-glue_pairs.json +14 -0
- wisent/examples/scripts/results/test_basque2/test_basque-glue_evaluation.json +51 -0
- wisent/examples/scripts/results/test_basque2/test_basque-glue_pairs.json +14 -0
- wisent/examples/scripts/results/test_basque_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_basque_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_basque_glue/test_basque-glue_evaluation.json +51 -0
- wisent/examples/scripts/results/test_basque_glue/test_basque-glue_pairs.json +14 -0
- wisent/examples/scripts/results/test_basqueglue_evaluation.json +51 -0
- wisent/examples/scripts/results/test_basqueglue_pairs.json +14 -0
- wisent/examples/scripts/results/test_bbh_evaluation.json +51 -0
- wisent/examples/scripts/results/test_bbh_pairs.json +14 -0
- wisent/examples/scripts/results/test_bbq_evaluation.json +30 -0
- wisent/examples/scripts/results/test_bbq_pairs.json +8 -0
- wisent/examples/scripts/results/test_bec2016eu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_bec2016eu_pairs.json +14 -0
- wisent/examples/scripts/results/test_belebele_evaluation.json +51 -0
- wisent/examples/scripts/results/test_belebele_pairs.json +14 -0
- wisent/examples/scripts/results/test_benchmarks_evaluation.json +51 -0
- wisent/examples/scripts/results/test_benchmarks_pairs.json +14 -0
- wisent/examples/scripts/results/test_bertaqa_evaluation.json +51 -0
- wisent/examples/scripts/results/test_bertaqa_pairs.json +14 -0
- wisent/examples/scripts/results/test_bhtc_v2_evaluation.json +30 -0
- wisent/examples/scripts/results/test_bhtc_v2_pairs.json +8 -0
- wisent/examples/scripts/results/test_bigbench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_bigbench_pairs.json +14 -0
- wisent/examples/scripts/results/test_blimp_evaluation.json +51 -0
- wisent/examples/scripts/results/test_blimp_pairs.json +14 -0
- wisent/examples/scripts/results/test_boolq/test_boolq_evaluation.json +30 -0
- wisent/examples/scripts/results/test_boolq/test_boolq_pairs.json +8 -0
- wisent/examples/scripts/results/test_boolq-seq2seq_evaluation.json +30 -0
- wisent/examples/scripts/results/test_boolq-seq2seq_pairs.json +8 -0
- wisent/examples/scripts/results/test_boolq_evaluation.json +30 -0
- wisent/examples/scripts/results/test_boolq_pairs.json +8 -0
- wisent/examples/scripts/results/test_c4_evaluation.json +30 -0
- wisent/examples/scripts/results/test_c4_pairs.json +8 -0
- wisent/examples/scripts/results/test_cabreu_evaluation.json +30 -0
- wisent/examples/scripts/results/test_cabreu_pairs.json +8 -0
- wisent/examples/scripts/results/test_careqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_careqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_catalan_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_catalan_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_catalanqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_catalanqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_catcola_evaluation.json +30 -0
- wisent/examples/scripts/results/test_catcola_pairs.json +8 -0
- wisent/examples/scripts/results/test_cb_evaluation.json +30 -0
- wisent/examples/scripts/results/test_cb_pairs.json +8 -0
- wisent/examples/scripts/results/test_ceval/test_ceval_evaluation.json +51 -0
- wisent/examples/scripts/results/test_ceval/test_ceval_pairs.json +14 -0
- wisent/examples/scripts/results/test_ceval_accountant/test_ceval-valid_accountant_evaluation.json +30 -0
- wisent/examples/scripts/results/test_ceval_accountant/test_ceval-valid_accountant_pairs.json +8 -0
- wisent/examples/scripts/results/test_ceval_evaluation.json +51 -0
- wisent/examples/scripts/results/test_ceval_pairs.json +14 -0
- wisent/examples/scripts/results/test_ceval_valid/test_ceval_valid_evaluation.json +51 -0
- wisent/examples/scripts/results/test_ceval_valid/test_ceval_valid_pairs.json +14 -0
- wisent/examples/scripts/results/test_chain_of_thought_evaluation.json +51 -0
- wisent/examples/scripts/results/test_chain_of_thought_pairs.json +14 -0
- wisent/examples/scripts/results/test_chartqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_chartqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_claim_stance_topic_evaluation.json +30 -0
- wisent/examples/scripts/results/test_claim_stance_topic_pairs.json +8 -0
- wisent/examples/scripts/results/test_cmmlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_cmmlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_cnn_dailymail_evaluation.json +30 -0
- wisent/examples/scripts/results/test_cnn_dailymail_pairs.json +8 -0
- wisent/examples/scripts/results/test_cocoteros_es_evaluation.json +30 -0
- wisent/examples/scripts/results/test_cocoteros_es_pairs.json +8 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_go_evaluation.json +30 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_go_pairs.json +8 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_java_evaluation.json +30 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_java_pairs.json +8 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_javascript_evaluation.json +30 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_javascript_pairs.json +8 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_php_evaluation.json +30 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_php_pairs.json +8 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_python_evaluation.json +30 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_python_pairs.json +8 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_ruby_evaluation.json +30 -0
- wisent/examples/scripts/results/test_codexglue_code_to_text_ruby_pairs.json +8 -0
- wisent/examples/scripts/results/test_coedit_gec_evaluation.json +30 -0
- wisent/examples/scripts/results/test_coedit_gec_pairs.json +8 -0
- wisent/examples/scripts/results/test_cola_evaluation.json +30 -0
- wisent/examples/scripts/results/test_cola_pairs.json +8 -0
- wisent/examples/scripts/results/test_commonsense_qa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_commonsense_qa_pairs.json +8 -0
- wisent/examples/scripts/results/test_conala_evaluation.json +30 -0
- wisent/examples/scripts/results/test_conala_pairs.json +8 -0
- wisent/examples/scripts/results/test_concode_evaluation.json +30 -0
- wisent/examples/scripts/results/test_concode_pairs.json +8 -0
- wisent/examples/scripts/results/test_copa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_copa_pairs.json +8 -0
- wisent/examples/scripts/results/test_copal_id_evaluation.json +30 -0
- wisent/examples/scripts/results/test_copal_id_pairs.json +8 -0
- wisent/examples/scripts/results/test_coqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_coqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_coqcat_evaluation.json +30 -0
- wisent/examples/scripts/results/test_coqcat_pairs.json +8 -0
- wisent/examples/scripts/results/test_crows_pairs_evaluation.json +51 -0
- wisent/examples/scripts/results/test_crows_pairs_pairs.json +14 -0
- wisent/examples/scripts/results/test_csatqa_evaluation.json +51 -0
- wisent/examples/scripts/results/test_csatqa_pairs.json +14 -0
- wisent/examples/scripts/results/test_cycle_letters_evaluation.json +30 -0
- wisent/examples/scripts/results/test_cycle_letters_pairs.json +8 -0
- wisent/examples/scripts/results/test_darija_bench/test_darija_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_darija_bench/test_darija_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_darija_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_darija_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_darijahellaswag_evaluation.json +30 -0
- wisent/examples/scripts/results/test_darijahellaswag_pairs.json +8 -0
- wisent/examples/scripts/results/test_darijammlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_darijammlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_dbpedia_14_evaluation.json +30 -0
- wisent/examples/scripts/results/test_dbpedia_14_pairs.json +8 -0
- wisent/examples/scripts/results/test_drop_evaluation.json +30 -0
- wisent/examples/scripts/results/test_drop_pairs.json +8 -0
- wisent/examples/scripts/results/test_ds1000_evaluation.json +30 -0
- wisent/examples/scripts/results/test_ds1000_pairs.json +8 -0
- wisent/examples/scripts/results/test_egyhellaswag_evaluation.json +30 -0
- wisent/examples/scripts/results/test_egyhellaswag_pairs.json +8 -0
- wisent/examples/scripts/results/test_egymmlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_egymmlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_epec_koref_bin_evaluation.json +30 -0
- wisent/examples/scripts/results/test_epec_koref_bin_pairs.json +8 -0
- wisent/examples/scripts/results/test_eq_bench_evaluation.json +30 -0
- wisent/examples/scripts/results/test_eq_bench_pairs.json +8 -0
- wisent/examples/scripts/results/test_escola_evaluation.json +30 -0
- wisent/examples/scripts/results/test_escola_pairs.json +8 -0
- wisent/examples/scripts/results/test_ethics_cm_evaluation.json +30 -0
- wisent/examples/scripts/results/test_ethics_cm_pairs.json +8 -0
- wisent/examples/scripts/results/test_ethos_binary_evaluation.json +30 -0
- wisent/examples/scripts/results/test_ethos_binary_pairs.json +8 -0
- wisent/examples/scripts/results/test_eus_exams/test_eus_exams_evaluation.json +51 -0
- wisent/examples/scripts/results/test_eus_exams/test_eus_exams_pairs.json +14 -0
- wisent/examples/scripts/results/test_eus_exams_es_evaluation.json +51 -0
- wisent/examples/scripts/results/test_eus_exams_es_pairs.json +14 -0
- wisent/examples/scripts/results/test_eus_exams_evaluation.json +51 -0
- wisent/examples/scripts/results/test_eus_exams_pairs.json +14 -0
- wisent/examples/scripts/results/test_eus_proficiency_evaluation.json +30 -0
- wisent/examples/scripts/results/test_eus_proficiency_pairs.json +8 -0
- wisent/examples/scripts/results/test_eus_reading_evaluation.json +30 -0
- wisent/examples/scripts/results/test_eus_reading_pairs.json +8 -0
- wisent/examples/scripts/results/test_eus_trivia_evaluation.json +30 -0
- wisent/examples/scripts/results/test_eus_trivia_pairs.json +8 -0
- wisent/examples/scripts/results/test_evalita-mp_evaluation.json +51 -0
- wisent/examples/scripts/results/test_evalita-mp_pairs.json +14 -0
- wisent/examples/scripts/results/test_evalita-sp_sum_task_fp-small_p1_evaluation.json +30 -0
- wisent/examples/scripts/results/test_evalita-sp_sum_task_fp-small_p1_pairs.json +8 -0
- wisent/examples/scripts/results/test_evalita_LLM_evaluation.json +51 -0
- wisent/examples/scripts/results/test_evalita_LLM_pairs.json +14 -0
- wisent/examples/scripts/results/test_evalita_llm/test_evalita_llm_evaluation.json +51 -0
- wisent/examples/scripts/results/test_evalita_llm/test_evalita_llm_pairs.json +14 -0
- wisent/examples/scripts/results/test_evalita_mp/test_evalita-mp_te_prompt-1_evaluation.json +30 -0
- wisent/examples/scripts/results/test_evalita_mp/test_evalita-mp_te_prompt-1_pairs.json +8 -0
- wisent/examples/scripts/results/test_evalita_mp2/test_evalita_mp_evaluation.json +51 -0
- wisent/examples/scripts/results/test_evalita_mp2/test_evalita_mp_pairs.json +14 -0
- wisent/examples/scripts/results/test_evalita_sp2/test_evalita-sp_sum_task_fp-small_p1_evaluation.json +30 -0
- wisent/examples/scripts/results/test_evalita_sp2/test_evalita-sp_sum_task_fp-small_p1_pairs.json +8 -0
- wisent/examples/scripts/results/test_fda_evaluation.json +30 -0
- wisent/examples/scripts/results/test_fda_pairs.json +8 -0
- wisent/examples/scripts/results/test_financial_tweets_evaluation.json +30 -0
- wisent/examples/scripts/results/test_financial_tweets_pairs.json +8 -0
- wisent/examples/scripts/results/test_fld/test_fld_evaluation.json +30 -0
- wisent/examples/scripts/results/test_fld/test_fld_pairs.json +8 -0
- wisent/examples/scripts/results/test_fld_evaluation.json +30 -0
- wisent/examples/scripts/results/test_fld_fixed/test_fld_evaluation.json +30 -0
- wisent/examples/scripts/results/test_fld_fixed/test_fld_pairs.json +8 -0
- wisent/examples/scripts/results/test_fld_pairs.json +8 -0
- wisent/examples/scripts/results/test_flores_evaluation.json +51 -0
- wisent/examples/scripts/results/test_flores_pairs.json +14 -0
- wisent/examples/scripts/results/test_freebase_evaluation.json +30 -0
- wisent/examples/scripts/results/test_freebase_pairs.json +8 -0
- wisent/examples/scripts/results/test_french_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_french_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_galcola_evaluation.json +30 -0
- wisent/examples/scripts/results/test_galcola_pairs.json +8 -0
- wisent/examples/scripts/results/test_galician_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_galician_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_glianorex_evaluation.json +30 -0
- wisent/examples/scripts/results/test_glianorex_pairs.json +8 -0
- wisent/examples/scripts/results/test_global_mmlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_global_mmlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_glue_evaluation.json +51 -0
- wisent/examples/scripts/results/test_glue_pairs.json +14 -0
- wisent/examples/scripts/results/test_gpqa_evaluation.json +51 -0
- wisent/examples/scripts/results/test_gpqa_pairs.json +14 -0
- wisent/examples/scripts/results/test_gpt3_translation_benchmarks_evaluation.json +51 -0
- wisent/examples/scripts/results/test_gpt3_translation_benchmarks_pairs.json +14 -0
- wisent/examples/scripts/results/test_groundcocoa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_groundcocoa_pairs.json +8 -0
- wisent/examples/scripts/results/test_gsm8k_evaluation.json +30 -0
- wisent/examples/scripts/results/test_gsm8k_pairs.json +8 -0
- wisent/examples/scripts/results/test_haerae_evaluation.json +51 -0
- wisent/examples/scripts/results/test_haerae_pairs.json +14 -0
- wisent/examples/scripts/results/test_headqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_headqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_hellaswag_evaluation.json +30 -0
- wisent/examples/scripts/results/test_hellaswag_pairs.json +8 -0
- wisent/examples/scripts/results/test_hendrycks_ethics_evaluation.json +51 -0
- wisent/examples/scripts/results/test_hendrycks_ethics_pairs.json +14 -0
- wisent/examples/scripts/results/test_hendrycks_math_evaluation.json +51 -0
- wisent/examples/scripts/results/test_hendrycks_math_pairs.json +14 -0
- wisent/examples/scripts/results/test_histoires_morales_evaluation.json +30 -0
- wisent/examples/scripts/results/test_histoires_morales_pairs.json +8 -0
- wisent/examples/scripts/results/test_hmmt_evaluation.json +30 -0
- wisent/examples/scripts/results/test_hmmt_feb_2025_evaluation.json +30 -0
- wisent/examples/scripts/results/test_hmmt_feb_2025_pairs.json +8 -0
- wisent/examples/scripts/results/test_hmmt_pairs.json +8 -0
- wisent/examples/scripts/results/test_hrm8k_evaluation.json +51 -0
- wisent/examples/scripts/results/test_hrm8k_pairs.json +14 -0
- wisent/examples/scripts/results/test_humaneval_evaluation.json +30 -0
- wisent/examples/scripts/results/test_humaneval_pairs.json +8 -0
- wisent/examples/scripts/results/test_humaneval_plus_evaluation.json +30 -0
- wisent/examples/scripts/results/test_humaneval_plus_pairs.json +8 -0
- wisent/examples/scripts/results/test_ifeval_evaluation.json +30 -0
- wisent/examples/scripts/results/test_ifeval_pairs.json +8 -0
- wisent/examples/scripts/results/test_instruct_humaneval/test_instruct_humaneval_evaluation.json +30 -0
- wisent/examples/scripts/results/test_instruct_humaneval/test_instruct_humaneval_pairs.json +8 -0
- wisent/examples/scripts/results/test_instruct_humaneval_evaluation.json +30 -0
- wisent/examples/scripts/results/test_instruct_humaneval_pairs.json +8 -0
- wisent/examples/scripts/results/test_inverse_scaling_evaluation.json +51 -0
- wisent/examples/scripts/results/test_inverse_scaling_hindsight_neglect_10shot_evaluation.json +30 -0
- wisent/examples/scripts/results/test_inverse_scaling_hindsight_neglect_10shot_pairs.json +8 -0
- wisent/examples/scripts/results/test_inverse_scaling_mc/test_inverse_scaling_mc_evaluation.json +51 -0
- wisent/examples/scripts/results/test_inverse_scaling_mc/test_inverse_scaling_mc_pairs.json +14 -0
- wisent/examples/scripts/results/test_inverse_scaling_pairs.json +14 -0
- wisent/examples/scripts/results/test_iwslt2017-ar-en_evaluation.json +30 -0
- wisent/examples/scripts/results/test_iwslt2017-ar-en_pairs.json +8 -0
- wisent/examples/scripts/results/test_iwslt2017-en-ar_evaluation.json +30 -0
- wisent/examples/scripts/results/test_iwslt2017-en-ar_pairs.json +8 -0
- wisent/examples/scripts/results/test_iwslt2017_ar_en/test_iwslt2017-ar-en_evaluation.json +30 -0
- wisent/examples/scripts/results/test_iwslt2017_ar_en/test_iwslt2017-ar-en_pairs.json +8 -0
- wisent/examples/scripts/results/test_iwslt2017_en_ar/test_iwslt2017-en-ar_evaluation.json +30 -0
- wisent/examples/scripts/results/test_iwslt2017_en_ar/test_iwslt2017-en-ar_pairs.json +8 -0
- wisent/examples/scripts/results/test_iwslt2017_group/test_iwslt2017_evaluation.json +30 -0
- wisent/examples/scripts/results/test_iwslt2017_group/test_iwslt2017_pairs.json +8 -0
- wisent/examples/scripts/results/test_japanese_leaderboard_evaluation.json +51 -0
- wisent/examples/scripts/results/test_japanese_leaderboard_pairs.json +14 -0
- wisent/examples/scripts/results/test_jsonschema_bench/test_jsonschema_bench_evaluation.json +30 -0
- wisent/examples/scripts/results/test_jsonschema_bench/test_jsonschema_bench_pairs.json +8 -0
- wisent/examples/scripts/results/test_jsonschema_bench_evaluation.json +30 -0
- wisent/examples/scripts/results/test_jsonschema_bench_final/test_jsonschema_bench_evaluation.json +30 -0
- wisent/examples/scripts/results/test_jsonschema_bench_final/test_jsonschema_bench_pairs.json +8 -0
- wisent/examples/scripts/results/test_jsonschema_bench_pairs.json +8 -0
- wisent/examples/scripts/results/test_kbl_evaluation.json +51 -0
- wisent/examples/scripts/results/test_kbl_fixed/test_kbl_evaluation.json +51 -0
- wisent/examples/scripts/results/test_kbl_fixed/test_kbl_pairs.json +14 -0
- wisent/examples/scripts/results/test_kbl_pairs.json +14 -0
- wisent/examples/scripts/results/test_kmmlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_kmmlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_kobest_evaluation.json +51 -0
- wisent/examples/scripts/results/test_kobest_pairs.json +14 -0
- wisent/examples/scripts/results/test_kormedmcqa/test_kormedmcqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_kormedmcqa/test_kormedmcqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_kormedmcqa_dentist/test_kormedmcqa_dentist_evaluation.json +30 -0
- wisent/examples/scripts/results/test_kormedmcqa_dentist/test_kormedmcqa_dentist_pairs.json +8 -0
- wisent/examples/scripts/results/test_kormedmcqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_kormedmcqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_lambada_cloze_evaluation.json +30 -0
- wisent/examples/scripts/results/test_lambada_cloze_pairs.json +8 -0
- wisent/examples/scripts/results/test_lambada_evaluation.json +30 -0
- wisent/examples/scripts/results/test_lambada_final/test_lambada_openai_mt_stablelm_en_evaluation.json +30 -0
- wisent/examples/scripts/results/test_lambada_final/test_lambada_openai_mt_stablelm_en_pairs.json +8 -0
- wisent/examples/scripts/results/test_lambada_multilingual/test_lambada_multilingual_evaluation.json +51 -0
- wisent/examples/scripts/results/test_lambada_multilingual/test_lambada_multilingual_pairs.json +14 -0
- wisent/examples/scripts/results/test_lambada_multilingual_evaluation.json +51 -0
- wisent/examples/scripts/results/test_lambada_multilingual_pairs.json +14 -0
- wisent/examples/scripts/results/test_lambada_multilingual_stablelm_evaluation.json +51 -0
- wisent/examples/scripts/results/test_lambada_multilingual_stablelm_pairs.json +14 -0
- wisent/examples/scripts/results/test_lambada_openai_evaluation.json +30 -0
- wisent/examples/scripts/results/test_lambada_openai_pairs.json +8 -0
- wisent/examples/scripts/results/test_lambada_pairs.json +8 -0
- wisent/examples/scripts/results/test_lambada_stablelm_en_fixed/test_lambada_openai_mt_stablelm_en_evaluation.json +30 -0
- wisent/examples/scripts/results/test_lambada_stablelm_en_fixed/test_lambada_openai_mt_stablelm_en_pairs.json +8 -0
- wisent/examples/scripts/results/test_lambada_stablelm_fixed/test_lambada_openai_mt_stablelm_en_evaluation.json +30 -0
- wisent/examples/scripts/results/test_lambada_stablelm_fixed/test_lambada_openai_mt_stablelm_en_pairs.json +8 -0
- wisent/examples/scripts/results/test_lambada_standard_evaluation.json +30 -0
- wisent/examples/scripts/results/test_lambada_standard_pairs.json +8 -0
- wisent/examples/scripts/results/test_leaderboard_evaluation.json +51 -0
- wisent/examples/scripts/results/test_leaderboard_pairs.json +14 -0
- wisent/examples/scripts/results/test_libra/test_libra_evaluation.json +51 -0
- wisent/examples/scripts/results/test_libra/test_libra_pairs.json +14 -0
- wisent/examples/scripts/results/test_libra_evaluation.json +51 -0
- wisent/examples/scripts/results/test_libra_pairs.json +14 -0
- wisent/examples/scripts/results/test_lingoly_evaluation.json +30 -0
- wisent/examples/scripts/results/test_lingoly_pairs.json +8 -0
- wisent/examples/scripts/results/test_livecodebench_evaluation.json +30 -0
- wisent/examples/scripts/results/test_livecodebench_pairs.json +8 -0
- wisent/examples/scripts/results/test_livemathbench_cnmo_en_evaluation.json +30 -0
- wisent/examples/scripts/results/test_livemathbench_cnmo_en_pairs.json +8 -0
- wisent/examples/scripts/results/test_livemathbench_cnmo_zh_evaluation.json +30 -0
- wisent/examples/scripts/results/test_livemathbench_cnmo_zh_pairs.json +8 -0
- wisent/examples/scripts/results/test_llama_evaluation.json +30 -0
- wisent/examples/scripts/results/test_llama_pairs.json +8 -0
- wisent/examples/scripts/results/test_logiqa2_evaluation.json +30 -0
- wisent/examples/scripts/results/test_logiqa2_pairs.json +8 -0
- wisent/examples/scripts/results/test_logiqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_logiqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_m_mmlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_m_mmlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_mastermind/test_mastermind_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mastermind/test_mastermind_pairs.json +14 -0
- wisent/examples/scripts/results/test_mastermind_24_easy/test_mastermind_24_easy_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mastermind_24_easy/test_mastermind_24_easy_pairs.json +8 -0
- wisent/examples/scripts/results/test_mastermind_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mastermind_pairs.json +14 -0
- wisent/examples/scripts/results/test_math500_evaluation.json +30 -0
- wisent/examples/scripts/results/test_math500_pairs.json +8 -0
- wisent/examples/scripts/results/test_math_evaluation.json +30 -0
- wisent/examples/scripts/results/test_math_pairs.json +8 -0
- wisent/examples/scripts/results/test_mathqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mathqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_mbpp_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mbpp_pairs.json +8 -0
- wisent/examples/scripts/results/test_mbpp_plus_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mbpp_plus_pairs.json +8 -0
- wisent/examples/scripts/results/test_mc_taco_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mc_taco_pairs.json +8 -0
- wisent/examples/scripts/results/test_med_concepts_qa/test_med_concepts_qa_evaluation.json +51 -0
- wisent/examples/scripts/results/test_med_concepts_qa/test_med_concepts_qa_pairs.json +14 -0
- wisent/examples/scripts/results/test_med_concepts_qa_atc_easy/test_med_concepts_qa_atc_easy_evaluation.json +30 -0
- wisent/examples/scripts/results/test_med_concepts_qa_atc_easy/test_med_concepts_qa_atc_easy_pairs.json +8 -0
- wisent/examples/scripts/results/test_med_concepts_qa_evaluation.json +51 -0
- wisent/examples/scripts/results/test_med_concepts_qa_pairs.json +14 -0
- wisent/examples/scripts/results/test_meddialog_evaluation.json +30 -0
- wisent/examples/scripts/results/test_meddialog_pairs.json +8 -0
- wisent/examples/scripts/results/test_meddialog_raw_perplexity/test_meddialog_raw_perplexity_evaluation.json +30 -0
- wisent/examples/scripts/results/test_meddialog_raw_perplexity/test_meddialog_raw_perplexity_pairs.json +8 -0
- wisent/examples/scripts/results/test_mediqa_qa2019_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mediqa_qa2019_pairs.json +8 -0
- wisent/examples/scripts/results/test_medmcqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_medmcqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_medqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_medqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_medtext_evaluation.json +30 -0
- wisent/examples/scripts/results/test_medtext_pairs.json +8 -0
- wisent/examples/scripts/results/test_mela_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mela_pairs.json +14 -0
- wisent/examples/scripts/results/test_meqsum_evaluation.json +30 -0
- wisent/examples/scripts/results/test_meqsum_pairs.json +8 -0
- wisent/examples/scripts/results/test_mercury_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mercury_pairs.json +8 -0
- wisent/examples/scripts/results/test_metabench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_metabench_pairs.json +14 -0
- wisent/examples/scripts/results/test_mgsm_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mgsm_pairs.json +14 -0
- wisent/examples/scripts/results/test_mimic_repsum_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mimic_repsum_pairs.json +8 -0
- wisent/examples/scripts/results/test_minerva_math_evaluation.json +51 -0
- wisent/examples/scripts/results/test_minerva_math_pairs.json +14 -0
- wisent/examples/scripts/results/test_mlqa_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mlqa_pairs.json +14 -0
- wisent/examples/scripts/results/test_mmlu-pro-plus_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mmlu-pro-plus_pairs.json +14 -0
- wisent/examples/scripts/results/test_mmlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mmlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_mmlu_pro_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mmlu_pro_pairs.json +14 -0
- wisent/examples/scripts/results/test_mmlu_prox_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mmlu_prox_pairs.json +14 -0
- wisent/examples/scripts/results/test_mmlusr_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mmlusr_pairs.json +8 -0
- wisent/examples/scripts/results/test_mmmu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_mmmu_pairs.json +14 -0
- wisent/examples/scripts/results/test_mnli_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mnli_pairs.json +8 -0
- wisent/examples/scripts/results/test_model_written_evals_evaluation.json +51 -0
- wisent/examples/scripts/results/test_model_written_evals_pairs.json +14 -0
- wisent/examples/scripts/results/test_moral_stories_evaluation.json +30 -0
- wisent/examples/scripts/results/test_moral_stories_pairs.json +8 -0
- wisent/examples/scripts/results/test_mts_dialog_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mts_dialog_pairs.json +8 -0
- wisent/examples/scripts/results/test_multiblimp_evaluation.json +51 -0
- wisent/examples/scripts/results/test_multiblimp_pairs.json +14 -0
- wisent/examples/scripts/results/test_multimedqa_evaluation.json +51 -0
- wisent/examples/scripts/results/test_multimedqa_pairs.json +14 -0
- wisent/examples/scripts/results/test_multipl_e_evaluation.json +30 -0
- wisent/examples/scripts/results/test_multipl_e_pairs.json +8 -0
- wisent/examples/scripts/results/test_mutual_evaluation.json +30 -0
- wisent/examples/scripts/results/test_mutual_pairs.json +8 -0
- wisent/examples/scripts/results/test_non_greedy_robustness_agieval_aqua_rat_evaluation.json +30 -0
- wisent/examples/scripts/results/test_non_greedy_robustness_agieval_aqua_rat_pairs.json +8 -0
- wisent/examples/scripts/results/test_noreval_evaluation.json +51 -0
- wisent/examples/scripts/results/test_noreval_pairs.json +14 -0
- wisent/examples/scripts/results/test_noticia_evaluation.json +30 -0
- wisent/examples/scripts/results/test_noticia_pairs.json +8 -0
- wisent/examples/scripts/results/test_nq_open_evaluation.json +30 -0
- wisent/examples/scripts/results/test_nq_open_pairs.json +8 -0
- wisent/examples/scripts/results/test_olaph_evaluation.json +30 -0
- wisent/examples/scripts/results/test_olaph_pairs.json +8 -0
- wisent/examples/scripts/results/test_openbookqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_openbookqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_openllm_evaluation.json +51 -0
- wisent/examples/scripts/results/test_openllm_pairs.json +14 -0
- wisent/examples/scripts/results/test_option_order_robustness_agieval_aqua_rat_evaluation.json +30 -0
- wisent/examples/scripts/results/test_option_order_robustness_agieval_aqua_rat_pairs.json +8 -0
- wisent/examples/scripts/results/test_paloma_evaluation.json +51 -0
- wisent/examples/scripts/results/test_paloma_pairs.json +14 -0
- wisent/examples/scripts/results/test_passkey/test_passkey_evaluation.json +30 -0
- wisent/examples/scripts/results/test_passkey/test_passkey_pairs.json +8 -0
- wisent/examples/scripts/results/test_paws-x_evaluation.json +51 -0
- wisent/examples/scripts/results/test_paws-x_pairs.json +14 -0
- wisent/examples/scripts/results/test_paws_en/test_paws_en_evaluation.json +30 -0
- wisent/examples/scripts/results/test_paws_en/test_paws_en_pairs.json +8 -0
- wisent/examples/scripts/results/test_penn_treebank_evaluation.json +30 -0
- wisent/examples/scripts/results/test_penn_treebank_pairs.json +8 -0
- wisent/examples/scripts/results/test_pile_10k/test_pile_10k_evaluation.json +30 -0
- wisent/examples/scripts/results/test_pile_10k/test_pile_10k_pairs.json +8 -0
- wisent/examples/scripts/results/test_piqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_piqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_polemo2_evaluation.json +30 -0
- wisent/examples/scripts/results/test_polemo2_pairs.json +8 -0
- wisent/examples/scripts/results/test_polymath_en_high_evaluation.json +30 -0
- wisent/examples/scripts/results/test_polymath_en_high_pairs.json +8 -0
- wisent/examples/scripts/results/test_polymath_en_medium_evaluation.json +30 -0
- wisent/examples/scripts/results/test_polymath_en_medium_pairs.json +8 -0
- wisent/examples/scripts/results/test_polymath_zh_high_evaluation.json +30 -0
- wisent/examples/scripts/results/test_polymath_zh_high_pairs.json +8 -0
- wisent/examples/scripts/results/test_polymath_zh_medium_evaluation.json +30 -0
- wisent/examples/scripts/results/test_polymath_zh_medium_pairs.json +8 -0
- wisent/examples/scripts/results/test_portuguese_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_portuguese_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat/test_prompt_robustness_agieval_aqua_rat_evaluation.json +30 -0
- wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat/test_prompt_robustness_agieval_aqua_rat_pairs.json +8 -0
- wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat_evaluation.json +30 -0
- wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat_pairs.json +8 -0
- wisent/examples/scripts/results/test_prost_evaluation.json +30 -0
- wisent/examples/scripts/results/test_prost_pairs.json +8 -0
- wisent/examples/scripts/results/test_ptb_evaluation.json +30 -0
- wisent/examples/scripts/results/test_ptb_pairs.json +8 -0
- wisent/examples/scripts/results/test_pubmedqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_pubmedqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_pythia_evaluation.json +51 -0
- wisent/examples/scripts/results/test_pythia_pairs.json +14 -0
- wisent/examples/scripts/results/test_qa4mre_evaluation.json +30 -0
- wisent/examples/scripts/results/test_qa4mre_pairs.json +8 -0
- wisent/examples/scripts/results/test_qasper_evaluation.json +30 -0
- wisent/examples/scripts/results/test_qasper_pairs.json +8 -0
- wisent/examples/scripts/results/test_race_evaluation.json +30 -0
- wisent/examples/scripts/results/test_race_pairs.json +8 -0
- wisent/examples/scripts/results/test_realtoxicityprompts_evaluation.json +30 -0
- wisent/examples/scripts/results/test_realtoxicityprompts_pairs.json +8 -0
- wisent/examples/scripts/results/test_recode_evaluation.json +30 -0
- wisent/examples/scripts/results/test_recode_pairs.json +8 -0
- wisent/examples/scripts/results/test_record_evaluation.json +30 -0
- wisent/examples/scripts/results/test_record_pairs.json +8 -0
- wisent/examples/scripts/results/test_ruler_evaluation.json +51 -0
- wisent/examples/scripts/results/test_ruler_pairs.json +14 -0
- wisent/examples/scripts/results/test_sciq_evaluation.json +30 -0
- wisent/examples/scripts/results/test_sciq_pairs.json +8 -0
- wisent/examples/scripts/results/test_score_evaluation.json +51 -0
- wisent/examples/scripts/results/test_score_pairs.json +14 -0
- wisent/examples/scripts/results/test_self_consistency_evaluation.json +30 -0
- wisent/examples/scripts/results/test_self_consistency_pairs.json +8 -0
- wisent/examples/scripts/results/test_siqa/test_siqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_siqa/test_siqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_siqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_siqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_spanish_bench_evaluation.json +51 -0
- wisent/examples/scripts/results/test_spanish_bench_pairs.json +14 -0
- wisent/examples/scripts/results/test_squad2_evaluation.json +30 -0
- wisent/examples/scripts/results/test_squad2_pairs.json +8 -0
- wisent/examples/scripts/results/test_squadv2_evaluation.json +30 -0
- wisent/examples/scripts/results/test_squadv2_pairs.json +8 -0
- wisent/examples/scripts/results/test_super-glue-lm-eval-v1-seq2seq_evaluation.json +30 -0
- wisent/examples/scripts/results/test_super-glue-lm-eval-v1-seq2seq_pairs.json +8 -0
- wisent/examples/scripts/results/test_super-glue-lm-eval-v1_evaluation.json +51 -0
- wisent/examples/scripts/results/test_super-glue-lm-eval-v1_pairs.json +14 -0
- wisent/examples/scripts/results/test_swag_evaluation.json +30 -0
- wisent/examples/scripts/results/test_swag_pairs.json +8 -0
- wisent/examples/scripts/results/test_tinyBenchmarks_evaluation.json +51 -0
- wisent/examples/scripts/results/test_tinyBenchmarks_pairs.json +14 -0
- wisent/examples/scripts/results/test_tmmluplus_evaluation.json +51 -0
- wisent/examples/scripts/results/test_tmmluplus_pairs.json +14 -0
- wisent/examples/scripts/results/test_translation_evaluation.json +51 -0
- wisent/examples/scripts/results/test_translation_pairs.json +14 -0
- wisent/examples/scripts/results/test_triviaqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_triviaqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_truthfulqa-multi_evaluation.json +51 -0
- wisent/examples/scripts/results/test_truthfulqa-multi_pairs.json +14 -0
- wisent/examples/scripts/results/test_truthfulqa_evaluation.json +30 -0
- wisent/examples/scripts/results/test_truthfulqa_mc1_evaluation.json +30 -0
- wisent/examples/scripts/results/test_truthfulqa_mc1_pairs.json +8 -0
- wisent/examples/scripts/results/test_truthfulqa_mc2_evaluation.json +30 -0
- wisent/examples/scripts/results/test_truthfulqa_mc2_pairs.json +8 -0
- wisent/examples/scripts/results/test_truthfulqa_pairs.json +8 -0
- wisent/examples/scripts/results/test_turkishmmlu_evaluation.json +51 -0
- wisent/examples/scripts/results/test_turkishmmlu_pairs.json +14 -0
- wisent/examples/scripts/results/test_unfair_tos_evaluation.json +30 -0
- wisent/examples/scripts/results/test_unfair_tos_pairs.json +8 -0
- wisent/examples/scripts/results/test_unscramble_evaluation.json +51 -0
- wisent/examples/scripts/results/test_unscramble_pairs.json +14 -0
- wisent/examples/scripts/results/test_webqs_evaluation.json +30 -0
- wisent/examples/scripts/results/test_webqs_pairs.json +8 -0
- wisent/examples/scripts/results/test_wikitext103_evaluation.json +30 -0
- wisent/examples/scripts/results/test_wikitext103_pairs.json +8 -0
- wisent/examples/scripts/results/test_wikitext_evaluation.json +30 -0
- wisent/examples/scripts/results/test_wikitext_pairs.json +8 -0
- wisent/examples/scripts/results/test_winogender_evaluation.json +51 -0
- wisent/examples/scripts/results/test_winogender_pairs.json +14 -0
- wisent/examples/scripts/results/test_winogrande_evaluation.json +30 -0
- wisent/examples/scripts/results/test_winogrande_pairs.json +8 -0
- wisent/examples/scripts/results/test_wmdp_evaluation.json +30 -0
- wisent/examples/scripts/results/test_wmdp_pairs.json +8 -0
- wisent/examples/scripts/results/test_wmt-ro-en-t5-prompt_evaluation.json +30 -0
- wisent/examples/scripts/results/test_wmt-ro-en-t5-prompt_pairs.json +8 -0
- wisent/examples/scripts/results/test_wmt14_en_fr_evaluation.json +30 -0
- wisent/examples/scripts/results/test_wmt14_en_fr_pairs.json +8 -0
- wisent/examples/scripts/results/test_wmt16_en_de_evaluation.json +30 -0
- wisent/examples/scripts/results/test_wmt16_en_de_pairs.json +8 -0
- wisent/examples/scripts/results/test_wmt16_ro_en_evaluation.json +30 -0
- wisent/examples/scripts/results/test_wmt16_ro_en_pairs.json +8 -0
- wisent/examples/scripts/results/test_wsc273_evaluation.json +30 -0
- wisent/examples/scripts/results/test_wsc273_pairs.json +8 -0
- wisent/examples/scripts/results/test_xcopa_evaluation.json +51 -0
- wisent/examples/scripts/results/test_xcopa_pairs.json +14 -0
- wisent/examples/scripts/results/test_xnli_eu_evaluation.json +30 -0
- wisent/examples/scripts/results/test_xnli_eu_pairs.json +8 -0
- wisent/examples/scripts/results/test_xnli_evaluation.json +51 -0
- wisent/examples/scripts/results/test_xnli_pairs.json +14 -0
- wisent/examples/scripts/results/test_xquad_evaluation.json +51 -0
- wisent/examples/scripts/results/test_xquad_pairs.json +14 -0
- wisent/examples/scripts/results/test_xstorycloze_evaluation.json +51 -0
- wisent/examples/scripts/results/test_xstorycloze_pairs.json +14 -0
- wisent/examples/scripts/results/test_xsum_evaluation.json +30 -0
- wisent/examples/scripts/results/test_xsum_pairs.json +8 -0
- wisent/examples/scripts/results/test_xwinograd_evaluation.json +51 -0
- wisent/examples/scripts/results/test_xwinograd_pairs.json +14 -0
- wisent/examples/scripts/results/test_yahoo_answers_topics_evaluation.json +30 -0
- wisent/examples/scripts/results/test_yahoo_answers_topics_pairs.json +8 -0
- wisent/parameters/__init__.py +1 -0
- wisent/parameters/lm_eval/all_lm_eval_task_families.json +169 -0
- wisent/parameters/lm_eval/broken_in_lm_eval.json +10 -0
- wisent/parameters/lm_eval/evaluations_not_lm_eval_tasks.json +0 -0
- wisent/parameters/lm_eval/evaluator_check.json +3476 -0
- wisent/parameters/lm_eval/final_verification.json +24782 -0
- wisent/parameters/lm_eval/group_task_evaluators.json +1833 -0
- wisent/parameters/lm_eval/group_tasks.json +150 -0
- wisent/parameters/lm_eval/individual_tasks.json +402 -0
- wisent/parameters/lm_eval/no_readmes.json +1 -0
- wisent/parameters/lm_eval/not_lm_eval_tasks.json +110 -0
- wisent/parameters/lm_eval/read_tasks.json +208 -0
- wisent/parameters/lm_eval/readme_files.json +208 -0
- wisent/parameters/lm_eval/track_progress_not_lm_eval_tasks.json +128 -0
- wisent/parameters/tasks/missing_task_families.json +2963 -0
- wisent/parameters/tasks/remaining_tasks_to_implement.json +199 -0
- wisent/parameters/tasks/risks.json +10 -0
- wisent/parameters/tasks/skills.json +14 -0
- wisent/parameters/tasks/tasks.json +56031 -0
- wisent/scripts/run_quality_metrics_sweep.sh +315 -0
- wisent/tests/__init__.py +0 -0
- wisent/tests/examples/__init__.py +0 -0
- wisent/tests/examples/cli/__init__.py +0 -0
- wisent/tests/examples/cli/activations/__init__.py +0 -0
- wisent/tests/examples/cli/activations/test_get_activations.py +127 -0
- wisent/tests/examples/cli/classifier/__init__.py +0 -0
- wisent/tests/examples/cli/classifier/test_classifier_examples.py +141 -0
- wisent/tests/examples/cli/contrastive_pairs/__init__.py +0 -0
- wisent/tests/examples/cli/contrastive_pairs/test_generate_pairs.py +89 -0
- wisent/tests/examples/cli/evaluation/__init__.py +0 -0
- wisent/tests/examples/cli/evaluation/test_evaluation_examples.py +117 -0
- wisent/tests/examples/cli/generate/__init__.py +0 -0
- wisent/tests/examples/cli/generate/test_generate_with_classifier.py +146 -0
- wisent/tests/examples/cli/generate/test_generate_with_steering.py +149 -0
- wisent/tests/examples/cli/generate/test_only_generate.py +110 -0
- wisent/tests/examples/cli/multi_steering/__init__.py +0 -0
- wisent/tests/examples/cli/multi_steering/test_multi_steer_from_trained_vectors.py +210 -0
- wisent/tests/examples/cli/multi_steering/test_multi_steer_with_different_parameters.py +205 -0
- wisent/tests/examples/cli/multi_steering/test_train_and_multi_steer.py +174 -0
- wisent/tests/examples/cli/optimizer/__init__.py +0 -0
- wisent/tests/examples/cli/optimizer/test_optimize_sample_size.py +102 -0
- wisent/tests/examples/cli/optimizer/test_optimizer_examples.py +59 -0
- wisent/tests/examples/cli/steering/__init__.py +0 -0
- wisent/tests/examples/cli/steering/test_create_steering_vectors.py +135 -0
- wisent/tests/examples/cli/synthetic/__init__.py +0 -0
- wisent/tests/examples/cli/synthetic/test_synthetic_pairs.py +45 -0
- wisent/tests/nosense/__init__.py +6 -0
- wisent/tests/nosense/base_nosense.py +81 -0
- wisent/tests/nosense/math500_nosense.py +72 -0
- wisent/tests/nosense/test_robustness.py +336 -0
- wisent/tests/test_all_cli_commands.py +674 -0
- wisent/tests/test_geometry_comprehensive.py +327 -0
- wisent/tests/test_titan_geometry.py +257 -0
- wisent/tests/visualize_geometry.py +148 -0
- wisent-0.7.379.dist-info/METADATA +64 -0
- wisent-0.7.379.dist-info/RECORD +1720 -0
- wisent-0.7.379.dist-info/WHEEL +5 -0
- wisent-0.7.379.dist-info/entry_points.txt +2 -0
- wisent-0.7.379.dist-info/licenses/LICENSE +21 -0
- wisent-0.7.379.dist-info/top_level.txt +1 -0
|
@@ -0,0 +1,1720 @@
|
|
|
1
|
+
wisent/__init__.py,sha256=jXMUyFIIcSsWnNECcJrbJnDD33bboO3fY9GF9q8bKTU,1229
|
|
2
|
+
wisent/cli.py,sha256=XKzGIGstr38EowHYpr821c6YuV9Eaw3I1I3NvLztTO0,3960
|
|
3
|
+
wisent/core/__init__.py,sha256=x1MX4vKpKP3c2FuIHcFly-UkoZwGVnRPbzcFaxr_Jdo,1340
|
|
4
|
+
wisent/core/autonomous_agent.py,sha256=n0j2YSmhAfLt2WBaKM4o33SNQiiLYjvIRqfnT1nbO4M,48681
|
|
5
|
+
wisent/core/benchmark_extractors.py,sha256=BtGoL23n9EzSGoywymSh8uIhlr8T_LU8y-hp7et1vIk,13616
|
|
6
|
+
wisent/core/benchmark_registry.py,sha256=KQWMDEKNLBRvzd91HyH8SlhwIKXwpJtTWoSPcIuTn5Y,4627
|
|
7
|
+
wisent/core/bigcode_extractors.py,sha256=xKrJLdglcpykCv9TR72PVwu4cq3Fp5W6XuMYrdDHsVc,696
|
|
8
|
+
wisent/core/bigcode_integration.py,sha256=rP5YI9P0jgf2Zjooi8ANZtXzi0cxHBJveGa5XIZA23g,31544
|
|
9
|
+
wisent/core/branding.py,sha256=TBXNjPO_wXn8pReeWdSEqO8t1crPrCz1bv8pZHeOlog,3903
|
|
10
|
+
wisent/core/cli_logger.py,sha256=x3cZJLLGDwjr9R5sNQ3xCOeS6u2dKBNJGDDl0V9bIVA,768
|
|
11
|
+
wisent/core/config_manager.py,sha256=RfB9PT3K1Ns5EsQiSGQ1DSxdlG-cVMdaaUuqQMdCNOY,63560
|
|
12
|
+
wisent/core/detection_handling.py,sha256=L14US_y4UWaSabeAEzWn6tPM79qQFgwpO-BXvBBO7fA,11183
|
|
13
|
+
wisent/core/diversity_processors.py,sha256=wm4rFK8qrvYLcDpaJVHe-r6QtDYDRGty-Eyd0nXzcdM,6883
|
|
14
|
+
wisent/core/download_full_benchmarks.py,sha256=6JVmof4xKC--Lkj8r6v70KGRxPVTqu7YC3nVsb5m42g,60722
|
|
15
|
+
wisent/core/hyperparameter_optimizer.py,sha256=OHkk_uP91QSeysAyzCFLzGqJdbfH3cIUxJT1kWA-PpY,22987
|
|
16
|
+
wisent/core/layer.py,sha256=on_jCGb0CUhn8dpbY9ZbH3yXk1QoNqR6uUPsI6eX1hU,391
|
|
17
|
+
wisent/core/lm_eval_harness_ground_truth.py,sha256=pKya5ea5Lfn5wriNSknmqo7cWQY9bezYlH15UopRAV4,66730
|
|
18
|
+
wisent/core/main.py,sha256=WSrdXr_Jq9KxV-dZIMg4GZitr9IotWbZlxF2keOgsRs,4191
|
|
19
|
+
wisent/core/managed_cached_benchmarks.py,sha256=_q26-13t_GLFvkMA3M8tge0Sv6fFJmH_AJqCVl-Dt3w,23499
|
|
20
|
+
wisent/core/mixed_benchmark_sampler.py,sha256=J1R7Q7LOXE28P2Bydx_NrgblJv35rqdkCPkvZiAxp_c,13715
|
|
21
|
+
wisent/core/model_persistence.py,sha256=Pr1A3E0MxqwICKsFXmjKVq9KgG8Ip5UbtG-gujGFShc,9870
|
|
22
|
+
wisent/core/multi_steering.py,sha256=7RKYLBBFxEsrSc83KPakAU3WzvVx5XO88tShM2cLKZE,14921
|
|
23
|
+
wisent/core/parser.py,sha256=v99xCTbiyy2f4JzaVjz7bUtvn-uZEAj4S-u2OPikE24,68987
|
|
24
|
+
wisent/core/representation.py,sha256=hBl_N9qbr5Gsa7GCQ0nMWRm82RqYEfhd9cyf0PPH5LY,195
|
|
25
|
+
wisent/core/save_results.py,sha256=NJ7elQzbh3Q_ZefdTV8L-9QyutteWSBWfgTw78OBm1k,13725
|
|
26
|
+
wisent/core/steering.py,sha256=oiDIvHVm-n0eCS_Er9kv-Bu0H2JOqns87QhY5Oz-T2A,22790
|
|
27
|
+
wisent/core/steering_method.py,sha256=Y09jIImhZnRuHkbpeAfJT1hDovXi0bacqt3X3TVSo68,411
|
|
28
|
+
wisent/core/steering_optimizer.py,sha256=OdgOCLnEYUnjAiP19f_I80167K8W08CrScVhYyDlvuw,65587
|
|
29
|
+
wisent/core/task_interface.py,sha256=4xm6nrduFdXSZpJccahfZWtpI4rZvCQcsS7GCJWpbMU,4906
|
|
30
|
+
wisent/core/task_selector.py,sha256=JLMQ8E1HAvs4H4UqJRIvvWaktJsSAo3uz8oY135qO9g,7774
|
|
31
|
+
wisent/core/time_estimator.py,sha256=Frwwbg8DRaeFlxBNewhLPhveysfOZQDR7oMhcJO6kt0,5963
|
|
32
|
+
wisent/core/timing_calibration.py,sha256=zIu1lgIEX_lBAmVfHakOHSEOvUAtFI5y4i3sk5euuhM,6939
|
|
33
|
+
wisent/core/universal_subspace.py,sha256=D35PgZE42IwzjdyrTjmPcYou9YKJkbVZPfwCHh2jF1g,30728
|
|
34
|
+
wisent/core/user_model_config.py,sha256=FGzHYo1kD7qDPhpffvDkigV-z7aiLxHAtqiMEPf_sIM,6729
|
|
35
|
+
wisent/core/wisent.py,sha256=yglyOL-10ptlTVTshgjuXle3PXo-fdWJiZPx8oLyabM,21076
|
|
36
|
+
wisent/core/activations/__init__.py,sha256=ZT3aU-eU5-RR_QDkmKZRweCltJTWVof8mgbL3ONAXKw,839
|
|
37
|
+
wisent/core/activations/activations.py,sha256=XeExqWcnVNy6rVIl2aPp3ZVJ4BaPPKdQgrC6kCZXy4Y,4055
|
|
38
|
+
wisent/core/activations/activations_collector.py,sha256=u3dUOKr8g7ZLr-2DlH9Zn9xJARNUrVmxSbjwIe16qbE,21610
|
|
39
|
+
wisent/core/activations/prompt_construction_strategy.py,sha256=KpaAdj75D3t6JR47mx70vr82uCijsuaMIolIEXuciO8,1403
|
|
40
|
+
wisent/core/activations/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
41
|
+
wisent/core/activations/core/atoms.py,sha256=7NbDHuK9UUunMphrhBkTTwsniFPS0Y5_OAtJ9Kvr0ZE,9163
|
|
42
|
+
wisent/core/adapters/__init__.py,sha256=UZkssKYL4-Jl_F29x29MQr6L4899lpK-eFvFa_qGa8Y,712
|
|
43
|
+
wisent/core/adapters/audio.py,sha256=hqEXJ5NSfGstnQLmYkaBw_4TZ9Ft2qJaBMOKtObkb4Q,21712
|
|
44
|
+
wisent/core/adapters/base.py,sha256=ir75AwPA39sjpzGIs3aPFL2poZ34CVqVmygO-KZHbpA,12824
|
|
45
|
+
wisent/core/adapters/multimodal.py,sha256=-N_wM0Qw50C-PPjUE-H3jlyXSKaJ2FlL2aVUZnHldVI,25528
|
|
46
|
+
wisent/core/adapters/robotics.py,sha256=8rgf3zIYJMP_RPgyCJeIGjG-fNjXjkLz6Iww12mlUR4,21847
|
|
47
|
+
wisent/core/adapters/text.py,sha256=FRyw-VawAzUiLMFPsVIxVx60UeP6HUBQZOxFbflLu34,13832
|
|
48
|
+
wisent/core/adapters/video.py,sha256=M1p2rHFTNkN2LjteYXp3NL3L6v1E238XMR8Agv7O5wI,18461
|
|
49
|
+
wisent/core/agent/__init__.py,sha256=cAzpae7HWdNrpxCqANWUA2TKkqpm_hoSj8JAdbfmLwM,81
|
|
50
|
+
wisent/core/agent/budget.py,sha256=zBhMTvdVGLqM6vIhHzm5BO2dAp_ANuJvFPmZ8_9YQdM,24301
|
|
51
|
+
wisent/core/agent/device_benchmarks.py,sha256=2G_SuiM87gAS0TPZaMzS03yTOTHlbwJzpWNxYH60IJY,26206
|
|
52
|
+
wisent/core/agent/diagnose.py,sha256=8WOFmFI5nrXvo81egt3nPiYgufK1mGqQKaUDfFwWbZc,9324
|
|
53
|
+
wisent/core/agent/steer.py,sha256=4v3kkWTo4qerEJSa-D2CBUQE0ESQRN90Rd3ovL48uQo,9928
|
|
54
|
+
wisent/core/agent/timeout.py,sha256=WvYC9n9huf0b5daXX1vt2G4b3OEsU7irPd15nj_Twbc,4328
|
|
55
|
+
wisent/core/agent/diagnose/__init__.py,sha256=cAzpae7HWdNrpxCqANWUA2TKkqpm_hoSj8JAdbfmLwM,81
|
|
56
|
+
wisent/core/agent/diagnose/agent_classifier_decision.py,sha256=WrrIW-ig4V_OoVQUHhieENOhEMjy0t72Z-ElgxWQS5s,28978
|
|
57
|
+
wisent/core/agent/diagnose/classifier_marketplace.py,sha256=m02pwQkC4ROui6BhNhGJArdgBREI2xYuaRinm09VxM0,22113
|
|
58
|
+
wisent/core/agent/diagnose/create_classifier.py,sha256=IeyqObop2Z6jdEgFz-RIW37bqcOl2mUC2l_cFNdn8ww,44269
|
|
59
|
+
wisent/core/agent/diagnose/response_diagnostics.py,sha256=NT3jVAXMAE5nnS0bR2QXWMdX3MmugJhVKj1CrAp8wnI,10655
|
|
60
|
+
wisent/core/agent/diagnose/select_classifiers.py,sha256=XmmG3eAYkMNoPGE55-0gQTIGtDeqc_5sg93Mx03r__k,18353
|
|
61
|
+
wisent/core/agent/diagnose/synthetic_classifier_option.py,sha256=ZWsyZM7iUoFPlDtyWDjxi1qOeoVo9bUqFtl6eZ5pqzM,32703
|
|
62
|
+
wisent/core/agent/diagnose/tasks/__init__.py,sha256=IGnEhyeCWNagVjRpIDoX4Msb1rxfeWLrsPa3EvhiYeI,698
|
|
63
|
+
wisent/core/agent/diagnose/tasks/task_manager.py,sha256=XOWc-B-WBbmC17myulsvDs6E-z3Vv1qZ_jWaleMDuHw,62305
|
|
64
|
+
wisent/core/agent/diagnose/tasks/task_relevance.py,sha256=D4UBr0TqUNXkDZnNgA5wa4NYHSKtDaiugYeVg5zGQjs,3250
|
|
65
|
+
wisent/core/agent/diagnose/tasks/task_selector.py,sha256=ll34stireeqW-B_T4daf_91kujzVFQ8sOilk-JrxpHA,5414
|
|
66
|
+
wisent/core/classifier/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
|
|
67
|
+
wisent/core/classifier/models/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
|
|
68
|
+
wisent/core/classifiers/__init__.py,sha256=AZyqSkAxKUtmNFBHfFfK8DzU2UwelX6Xhl4VNLJ_4Yg,37
|
|
69
|
+
wisent/core/classifiers/classifiers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
70
|
+
wisent/core/classifiers/classifiers/rotator.py,sha256=nc3N-PuZwaBzbtk4n7cetZux1GB0BWPHTV7cACXP23I,5148
|
|
71
|
+
wisent/core/classifiers/classifiers/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
72
|
+
wisent/core/classifiers/classifiers/core/atoms.py,sha256=iy6oRuPUYJUN8UUpxu_FZ8wc9Yf-ZagFjTC82uZFGqU,26131
|
|
73
|
+
wisent/core/classifiers/classifiers/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
74
|
+
wisent/core/classifiers/classifiers/models/logistic.py,sha256=6FNnybW4zZ0HSeFjyEiOQrtNFN4LV-M39mmiiW4buDk,876
|
|
75
|
+
wisent/core/classifiers/classifiers/models/mlp.py,sha256=f_8Lgh5UZPLLCIKF1b6qeg5MwyFj9tuLrNMUcs-QDOo,1536
|
|
76
|
+
wisent/core/classifiers/core/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
|
|
77
|
+
wisent/core/classifiers/models/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
|
|
78
|
+
wisent/core/classifiers/pipeline_steps/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
|
|
79
|
+
wisent/core/cli/__init__.py,sha256=0paqfodiH1pMUIzqLbnc9HB-reukpjfwzmgmfUrOMX4,1929
|
|
80
|
+
wisent/core/cli/check_linearity.py,sha256=Ktu59KiXhTEL0U7pkD-H16xOdlhcs33dA08hkHSJXDc,4194
|
|
81
|
+
wisent/core/cli/create_steering_vector.py,sha256=zfWS9KPjQymKsBUbtndEE6jmOhuP4XMCYi1w82p98_4,16073
|
|
82
|
+
wisent/core/cli/diagnose_pairs.py,sha256=dopxJUgxien33yx6zAhwRVfb_4uvIem7Zpj2lawC1dQ,6775
|
|
83
|
+
wisent/core/cli/diagnose_vectors.py,sha256=jWLyk990j7jHtP1XVga6hVaR6LOzF6pzjwhkq--GZfs,17088
|
|
84
|
+
wisent/core/cli/estimate_unified_goodness_time.py,sha256=KnT7qWBkKwn7M1iaM62NRQpKdF_3pmoTQq4w7WYbRRU,16510
|
|
85
|
+
wisent/core/cli/evaluate_refusal.py,sha256=V9MjNEs4XqChnr1o9CrkgFwVVbcrhmNtl-p-0nkZ_jY,8758
|
|
86
|
+
wisent/core/cli/evaluate_responses.py,sha256=jDsb6qnnqJWivL80DERtvWN5nD7je68FNxVp3TNBBIA,38415
|
|
87
|
+
wisent/core/cli/generate_humanization_pairs.py,sha256=hspH8kd9u9BnI0Zdh7s2nMvCi45sMIvd9bLSIa9eLio,4200
|
|
88
|
+
wisent/core/cli/generate_pairs.py,sha256=KT88DaRMv5UcQwkLY9hf8uF3vuq9i3pWSnAL6Nejrzc,7545
|
|
89
|
+
wisent/core/cli/generate_pairs_from_task.py,sha256=YppLsiyDJhv4U7aoMBggLshLD5MsRQGirrLrg6oBLLc,4406
|
|
90
|
+
wisent/core/cli/generate_responses.py,sha256=nPEwQ8_EVjRbr0lhsCl8sacQzvt5XgDUOONSwRyI--U,5238
|
|
91
|
+
wisent/core/cli/generate_vector_from_synthetic.py,sha256=Fc9vcd7-1VTqHy8-gjTuPiCC7FIP1rxTajZf8TFSIi8,8868
|
|
92
|
+
wisent/core/cli/generate_vector_from_task.py,sha256=JMtlzUKeloNnbmivFdkwDN2i1IBFzicrFozIm3qeLz8,9837
|
|
93
|
+
wisent/core/cli/get_activations.py,sha256=ZXb9lNaVWuVLcJnDH40Eg_YRj8-tlLcx7Pxtk1dkqZ4,7778
|
|
94
|
+
wisent/core/cli/inference_config.py,sha256=iIZa0_ciBbM0Fqat0DJNd-mlWd0ANuYoeZ1v1V4jduc,3148
|
|
95
|
+
wisent/core/cli/inference_config_cli.py,sha256=CKvD3TpFYDSNtm9HdRFv0uMGInaR6iSHP8aLWYPhrfs,1786
|
|
96
|
+
wisent/core/cli/modify_weights.py,sha256=JhZwUl91lWsZwdiie60FQCxAfcwewtICBXf8XOLlBdw,28562
|
|
97
|
+
wisent/core/cli/multi_steer.py,sha256=GRs0557YabcEwch7P75uWJadHf4Ssa8YE5nTMt8WTvs,4564
|
|
98
|
+
wisent/core/cli/optimization_cache.py,sha256=nvnmFSy9km3tcquL36bhumCvc6wxl-_Ch3Ju9-MzNOE,9943
|
|
99
|
+
wisent/core/cli/optimize.py,sha256=WNjEa4cqpJCDk1xyBYNBEYkbdHC86lbuDyP_BIuhq3U,26730
|
|
100
|
+
wisent/core/cli/optimize_classification.py,sha256=2njWv90ZLT_yDxFgb6mNSZxOvcBsC5uMw5dZ7QGvKdw,23148
|
|
101
|
+
wisent/core/cli/optimize_sample_size.py,sha256=4HaJVQj3J-AZ-B2CURZ3wTddXPoPSrVIo8LKIxm3zQ8,16870
|
|
102
|
+
wisent/core/cli/optimize_steering.py,sha256=fxmm19bDB2UVmlpoBmnxpKPyH31zzLRAt_NIazm3A2Y,159088
|
|
103
|
+
wisent/core/cli/optimize_weights.py,sha256=5O6WMj7u7U--2DkXv_-MCT2uiZc8VEH4R6AoDFzoD3o,51731
|
|
104
|
+
wisent/core/cli/steering_method_trainer.py,sha256=q0UhaWW6ev3LcnU1pP_Lh0Na7D6gUNIB4OuIE4uAPyI,22912
|
|
105
|
+
wisent/core/cli/steering_search_space.py,sha256=tUcOohZT9_IUPFVnvkGHLyUjjovVGRNVRX4OCJOnaIY,22776
|
|
106
|
+
wisent/core/cli/tasks.py,sha256=pq0-K623l7qZ6hSxC1STbmpUi4i9qKshaOZ4-XEmeYE,44667
|
|
107
|
+
wisent/core/cli/train_unified_goodness.py,sha256=7T6vg6VeOejZvR488w2oOos_aWIBB7LWpK7fU4JSI_w,29353
|
|
108
|
+
wisent/core/cli/agent/__init__.py,sha256=tWH634YZVFfHmaEEQPgOn1LILIXwdpyJg5nj9tsdFz8,591
|
|
109
|
+
wisent/core/cli/agent/apply_steering.py,sha256=bWoAPUzUoSquuhj5N-fMtFsDzxJYmCT1aYthntbkB1k,7156
|
|
110
|
+
wisent/core/cli/agent/evaluate_response.py,sha256=VAmB2A7KvcgdzZkIG0XiYbjYUwAIbL3buYH-Hue-tJM,4901
|
|
111
|
+
wisent/core/cli/agent/generate_synthetic_pairs.py,sha256=QnMuJvunEB3Icmx9TrFeMklh1fjpTMLeu8mkV_dx46w,4723
|
|
112
|
+
wisent/core/cli/agent/main.py,sha256=aZ7SIuzTmYkFTaygjTf76Fgho9CiRMIOooes5Vd7ueI,5457
|
|
113
|
+
wisent/core/cli/agent/train_classifier.py,sha256=E8z3FXk2ZiWNtBZ8Q1o1SZ3XTag2ZajY7t9nXUsTQ0c,6934
|
|
114
|
+
wisent/core/contrastive_pairs/__init__.py,sha256=AbaAf-t_nyVVy_vLjp8WAlMDmNun3KNp_GMWAK25r9g,429
|
|
115
|
+
wisent/core/contrastive_pairs/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
116
|
+
wisent/core/contrastive_pairs/core/atoms.py,sha256=_zghw6c8iisW_SqBIUCoAnzhc5q7t5EgZ4zzTPxeLwQ,1129
|
|
117
|
+
wisent/core/contrastive_pairs/core/buliders.py,sha256=VWe4StOd3SsV0FBzg8b_2KG_ARiIITrwkfHR5xZNBxk,1975
|
|
118
|
+
wisent/core/contrastive_pairs/core/pair.py,sha256=eC-Zsp8GCrebjlNTxpkCq4G-OsuRTEZoJ2HJGVaClu0,8343
|
|
119
|
+
wisent/core/contrastive_pairs/core/response.py,sha256=JxElhJz3NCMkgP1ezbv0A2Zd-tA1iKwj4fqdvM-PiZY,6195
|
|
120
|
+
wisent/core/contrastive_pairs/core/serialization.py,sha256=mv0OD0kHiMknM1uH6sh1gyQeOcXFUIqJ2jBCgWh8dz0,11480
|
|
121
|
+
wisent/core/contrastive_pairs/core/set.py,sha256=sVJ2zvlVaFpMPgr4dYe-FMhT6sPS0KO94euksQN8aMo,6620
|
|
122
|
+
wisent/core/contrastive_pairs/diagnostics/__init__.py,sha256=X17XJthKEO4gUxUqBR01ysPwf7jx4LZoPKu3QZPK8Xk,2366
|
|
123
|
+
wisent/core/contrastive_pairs/diagnostics/activations.py,sha256=TUNxU-HV3oeQxFbZhKo-OISM4mzR-Bib0naHmbqWzk8,1736
|
|
124
|
+
wisent/core/contrastive_pairs/diagnostics/base.py,sha256=uBi8PdTd6BRyy0lmGjAZLTZdgiiWwPNtsmKkBFCmlD0,2658
|
|
125
|
+
wisent/core/contrastive_pairs/diagnostics/control_vectors.py,sha256=yLFtaXI9Fw2vjKpzDepCd6H_91g8G96V3xhToyHhzd0,63131
|
|
126
|
+
wisent/core/contrastive_pairs/diagnostics/coverage.py,sha256=MpT6_IdLqtMpav6mOCiNuemBVFvxWzkUbj7j3ZNx-48,2761
|
|
127
|
+
wisent/core/contrastive_pairs/diagnostics/divergence.py,sha256=Io3AcGluJogz4qENWu0ivQyFR_5bLN49BzCTI7DIVa4,3430
|
|
128
|
+
wisent/core/contrastive_pairs/diagnostics/duplicates.py,sha256=s5JPUdWcndX_kWAwP5UmXxEYNkW9DAYqtKwntIRoAiI,4755
|
|
129
|
+
wisent/core/contrastive_pairs/diagnostics/linearity.py,sha256=iN6pVQKUIvY-xbJSkzgJRULNxOy6dvWdvioTofH-zNk,11884
|
|
130
|
+
wisent/core/contrastive_pairs/diagnostics/vector_quality.py,sha256=YsR2Ii97Q5VOOQaX3wTSObL6cPVV1UiVl-JIj0Jv97c,22191
|
|
131
|
+
wisent/core/contrastive_pairs/huggingface_pairs/__init__.py,sha256=PYur_GOluHmwaYUS6rxsoas5-AmrG7OhVGoPteuJB-c,76
|
|
132
|
+
wisent/core/contrastive_pairs/huggingface_pairs/atoms.py,sha256=iIGJRHpDKLskBE_nNOVk9R5329OooktQ9r92_DZNq9A,8826
|
|
133
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_extractor_manifest.py,sha256=b3ofCWMNnYM1ijbjDdrzbA21o-_fhk-yVSr1QjaI894,27601
|
|
134
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_extractor_registry.py,sha256=OVD-qzWEzWX0vDz9j8oBiz2gK8rxnpOBZEC4_DOIfak,4680
|
|
135
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/__init__.py,sha256=SdkykJE5StRiI2OpXyXzOO-4jleibuBchWQZmGgRFDM,2542
|
|
136
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentbench.py,sha256=tdJznfmBU0PxtHwzgXVIJy5qRCGPQ898i_-Mu2FFoF8,7936
|
|
137
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentharm.py,sha256=hWEPf8mKIgZDZuisG0YMtAr04c33lfXjkRT8-wr2zfo,10504
|
|
138
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentic_search.py,sha256=fKSzaCY5Y2mRjRYr2OpP8RDQ6GN6LU3DHepq3rlNV30,17563
|
|
139
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aider_polyglot.py,sha256=u1KoeWTGdvO4jcxde56Jfrn_52O3JQPgxnThHgHTJ18,7335
|
|
140
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime.py,sha256=v0xmp650bG5NYHynaookiXrxstdZOU6zwYOhTtguPMI,3584
|
|
141
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime2024.py,sha256=iWtVKnwr1dS4k-zDU9nS4qsN3hnoFNjKb4JOIHsRuqM,2234
|
|
142
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime2025.py,sha256=l2NC4sA7yCpsEx-NtyQmj7x19KfXHcprsYhxVA8a6OQ,2132
|
|
143
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/alpaca_eval.py,sha256=NQWzHEcZr_kKDNcb4mXbpy2Rf45wOWEIxYR9o6YEs9s,5560
|
|
144
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/apps.py,sha256=WH5IXGOb5SB_IPXr1-_yVVb-OeMxR1wCntHOWsqUsUg,6297
|
|
145
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/arena_hard.py,sha256=T9qLzUpFLHZz1OBwGJ4_-7c6_3ZWswDebbvYle3sINw,6824
|
|
146
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/atis.py,sha256=UJfzAett3soThQGvnkFctninkNPMK5YFj16SVm0dsdw,3102
|
|
147
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/babilong.py,sha256=sqXHWyyRE1LMZB5IZtSdldxP7PP6OR890bZ68Vn4wt8,3050
|
|
148
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bangla_mmlu.py,sha256=JId_HcpSl8pn8nukmy4H2K11XLlD6XfUR-pn-YGffSI,3385
|
|
149
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/basqueglue.py,sha256=XMr4H1YUegf62AyC-KgTnNizpTXoN8gfmlRDAtYav6A,7666
|
|
150
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bec2016eu.py,sha256=jRg17Bd08mKGsbtC7xt6-5EfLEr93AUfk4578wHebS0,3806
|
|
151
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bfcl.py,sha256=LZmvmoJFGU9LUc5xxHbugjWiD__rADJkNrmIHJTC-nE,10536
|
|
152
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bhtc_v2.py,sha256=aVRRHAkLIHxpbSqYFXPwIYKdQLTo6X05_Rg3Dof4q_k,3080
|
|
153
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/browsecomp.py,sha256=g0Up4jGNzVf0g8SQtBdnCoZvNLTjLe52OV4hwOa7uto,9919
|
|
154
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/chain_of_thought.py,sha256=QsaJFZiFauHzlenUon7P0FnB9cDdMEVGDpJK2nm_w3g,3160
|
|
155
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/chinese_simpleqa.py,sha256=2q7r_PplOvbYUcNzHsk3JywEb5j6QjD9mR-Fyf7xw10,7107
|
|
156
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/cluewsc.py,sha256=yhi43rtNT_zjfZ2yg2BuxHI0kSDSETKc-tPk9eeFqgc,6300
|
|
157
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/cnn_dailymail.py,sha256=NSWDMh2ScUsvC_WVAgIWfGkkY64F_lPbbj9U4FokD7s,3200
|
|
158
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codeforces.py,sha256=TI5tRScQ4jqC_h8mV01fyZOddyR6Y4J0Vb8USBOB9yY,11774
|
|
159
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue.py,sha256=-Wk14hfIM0GOCjUilMniqoBz1CUOyZ4xwsRhzxuCZhA,3388
|
|
160
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text.py,sha256=5RH3NDPmrD7DYbt8LsYoEHExVaPOseFOTy_bd0Oh0pE,532
|
|
161
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_go.py,sha256=rAWrOb9-edJReFKFEp0orVdZzsBkUc1Jk0l652sMLPw,2142
|
|
162
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_java.py,sha256=2r1mQ-GVQpHrNJkTtr5jd1-s-c2DyrEifzNCPkIDSDA,2161
|
|
163
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_javascript.py,sha256=1qF8qRlIiQzinSPsV6DV6yHVb4mah1z4j_09_5CpxtM,2215
|
|
164
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_php.py,sha256=0ezAul2CAfxgOjS2zlc2lnsnO2ryB1YnBEyvLofJero,2152
|
|
165
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_python.py,sha256=vJ8V9WJofxXAdwict1IEjxYvVj0PsS7oK3khwpb9OXM,2179
|
|
166
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_ruby.py,sha256=nhprmMfG-82hi6DYlnvJrSYprgez4uBIUCS1Ta6WNhs,2161
|
|
167
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coding_benchmarks.py,sha256=QXJBlUj_rJIwXqn1uNouyT76hBgwOHMYJabaLIfE9ng,25867
|
|
168
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coedit_gec.py,sha256=a-0kJQFcAmoLNFE-dV-G07th9qXPQE0IuAXPcZaVEm8,2605
|
|
169
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/conala.py,sha256=auULgkRhG77vy2rAHJlpRhAsdTLrIZuq_XKK9PZF9tU,4075
|
|
170
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/concode.py,sha256=qhrgzJsJUnCG_hg4ouw_rMK5VFsecTwa4q7h_kXqMK8,3400
|
|
171
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/dbpedia_14.py,sha256=Ih4fFYMgPOrjYrxO9MaB2uG9W-xhK56B2tXVyA-dQFY,3123
|
|
172
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/doc_vqa.py,sha256=aLvaAU8ck2bC8GiETepKLHbG56mp7IESDJF9-z79Fw8,3726
|
|
173
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/donotanswer.py,sha256=OVf-VsgyAfEiy4mUXtTjdUG0ZQN6KCq0vggbhgaHkzo,8703
|
|
174
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ds1000.py,sha256=iJ09JUjTJkClELpRA8K6g2MbSgVcS2kfGWJJhoXlnnk,4318
|
|
175
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ds_1000.py,sha256=IjqjY3Bi38VIi0fAkXrcyLe_qsYLXVR13OJammtF1Uw,5022
|
|
176
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/epec_koref_bin.py,sha256=8JaVRVRfT1zawUnDecuHVWd3fOVzedvDSCZKKjnJzFo,2874
|
|
177
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ethos_binary.py,sha256=gkl6obtnHfJhwmUVL9MfTeVoscMQcBLL0bvQcebKU2k,2700
|
|
178
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/evalita_mp.py,sha256=Ioy59ZiNVnkwHV7IR0A6wJI3vFASr8o6xcXdrHjOT3I,7719
|
|
179
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/evalita_sp_sum_task_fp_small_p1.py,sha256=N-LunVyfI4jClY5svXq088V7N0QZR3aJqEPCBwmPjOg,3078
|
|
180
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/facts_grounding.py,sha256=Oyf65Wn2DzE4tSwvMkR6A4XAQLmmKqBNLvLuTgRApv0,6506
|
|
181
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/faithbench.py,sha256=qX_YGB9MBPfs_NhG7zfkxNTSiU-SFPbHfT1C01vkpMA,12124
|
|
182
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/financial_tweets.py,sha256=vUE79yRr5oE1vjFZ_wL4s6XFw7UuakWwrGf5UF24iAw,3635
|
|
183
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flames.py,sha256=wu-LUpq9irabMqMXfCHkfJui2Uyvj7LM0mqq_Jp_2yk,11430
|
|
184
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flan_held_in.py,sha256=Xzx2V6j8ihoYwC4SYG226SXXEH5j9G2vJFOMT0Y7pR8,3650
|
|
185
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flores.py,sha256=MwEv7rWS1_e9PMdmC88SM7yNkh189ZJgbhQXKMvmjT0,23149
|
|
186
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/frames.py,sha256=q2BY-zNDQ8pK_faY3JUxA7OUTfMX_kbOkILg-tjaYyc,5009
|
|
187
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/freebase.py,sha256=UqwnPzljWCSHOyaz7l4C2Zd6lJlVTXxvwy9gxVNeG5k,3801
|
|
188
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/get_negative_example_livecodebench.py,sha256=lrhwIQtQupD-xM9ypOamMcPuohkYHIb5uoDvmvgF8JA,5129
|
|
189
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/get_positive_example_livecodebench.py,sha256=k_PSaWtokvvBfysz3goFh-ZneXGSAKDaLQ22kkMl5oQ,4622
|
|
190
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/gpt3_translation_benchmarks.py,sha256=aOsJ-l67Noc1qnq_orp7-iw8wSNLX-VCa9C_gXX7MAc,3773
|
|
191
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hallucinations_leaderboard.py,sha256=ie_z2eq8XkY_ZzedZBMAoMgD0s1-TYUrwlNmpae7MOI,15385
|
|
192
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/halueval.py,sha256=N-nwmLP_Zsni0IvjgB2efPSGwJ0m95oW20eyx9Um0-I,8833
|
|
193
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/harmbench.py,sha256=HyHZYRFUSZFgJ7atINICsCCK3t01SofCaqbzUjdqrE0,9575
|
|
194
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/healthbench.py,sha256=_APNLxYsMgu3rJigXJwWdB83qL-X3pBMRJF0fzAPWNo,6831
|
|
195
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hle.py,sha256=awu6NcomLtATCwps67BQZJ0bZNDTx9EtYgoi9dshVlU,3176
|
|
196
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hmmt.py,sha256=oJ4UQvVHmdo1klTJlqJY3ckfbtD6wk3JYV-P8QAjfNY,3584
|
|
197
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/humaneval.py,sha256=CYjkdVaRQ3qyjdEFiKstZdmXNY0g-nt3_MP00gbHsNg,4220
|
|
198
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/humanevalpack.py,sha256=cVQOaOM7O_BpEYXnu8jqtpKpGBIG8BRe6MxyZ25qMHQ,3340
|
|
199
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/instruct_humaneval.py,sha256=Xgfr3rHmZ6K4E-0rY1u9af_nT5BTFzwHis7YKz7MUq4,6747
|
|
200
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/instructhumaneval.py,sha256=Xhr7sI4tqCmNyhzv3WrAM3fp17QJEe2oB2Nw6x9Od9Q,4712
|
|
201
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/iwslt2017_ar_en.py,sha256=OakHs-5Qli9tt9gXXsQ913u1aJqbNm4bP7hQ3kbAol0,3826
|
|
202
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/iwslt2017_en_ar.py,sha256=GszqSy52EgJ1L885KnhvA5zs_TEBbb6_NJP__aTXWgI,3826
|
|
203
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/jailbreakbench.py,sha256=DI9kig2nHMgLZ-Ow4EpwYTQlpNinTW1ul15ms_r4SJA,9642
|
|
204
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/law_stack_exchange.py,sha256=axujUVAn3ceAr6G3O4dfOZi2u9jxtboGlRerYdsUQYk,3577
|
|
205
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ledgar.py,sha256=fVsEeWfzTFG1yyEBbg5jq6y_vkMZcm2PbH7S4TXN_PU,4828
|
|
206
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench.py,sha256=gQjkxgF7qByJLOHN2JQwOgwY4IkJUWw7kESNyQlkT1g,1964
|
|
207
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench_contrastive_pair_generator.py,sha256=btT2P6utHmbejctjRelREYO_JCqV_Rhm4BWFJix11EQ,16808
|
|
208
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench_v6.py,sha256=4S6wv2yt6-s1mJeme-iEmTgXFHfOjBmpL4e6Y3QiseA,9486
|
|
209
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livemathbench.py,sha256=ugdpVXr7IP1S3YYsqzA0OQaXsYSW82OreUTMcK_QwhI,7792
|
|
210
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/llama.py,sha256=0GP8duCOy1u-G--s4VgIWNIihH06-0xSiE9GCqK82UA,3722
|
|
211
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/longform_writing.py,sha256=q44Q7zyqJUuWrKek70YXR34BPW5a5z9mojf5C1SIxxM,12728
|
|
212
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/m_mmlu.py,sha256=RDZmJAZfJdpxXwrbhDPFBOClNDV2gMf9TaBWNhVF3Gw,3724
|
|
213
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math.py,sha256=1HUMRBIN9b26aWOcyM70tEBbLfzIiZ9uw3OmKZFBpEA,5980
|
|
214
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math500.py,sha256=S5OlCvbvBkGN6-m8YvNXKZTSMnRM8I1bNF6MtnDSnYo,4760
|
|
215
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mbpp.py,sha256=iMiddPfOsMPhwH_y0RwRCAC0OVE4syCSmx9Q8jvHwPQ,4792
|
|
216
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/meddialog.py,sha256=n-Mh0iWhd-v6avJVAIKe3juHn_D5CrqsXCgERQNK8uA,2616
|
|
217
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medical_abstracts.py,sha256=TP1jcbcVRaDuwwSI_CusIwlchI2QWfMhsCa3rHLH_RY,3576
|
|
218
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medium_priority_benchmarks.py,sha256=-NHzHBPM8I0zVUkdCO4gsN0RyZ3vKJQHyP9RUNfKOso,30169
|
|
219
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mercury.py,sha256=_9EotHw83vgaUDxCnb0eS8TI5_hik68jlq9HdZy6lq4,3518
|
|
220
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mmlu_redux.py,sha256=VX9mKGZZtsaVggyXxfFKOL2pIs-V6Ah1UXN9rktsBCA,7611
|
|
221
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mmlusr.py,sha256=Ib5IPMeTp9m-Zc90YlxSXxJiZWo3tw8WasbuUL33yG0,3974
|
|
222
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multimedqa.py,sha256=0ijOZPjd86pZlUqDGHKK-QRJcJSSWPvbUglNy0uSw7o,3811
|
|
223
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multipl_e.py,sha256=Z_kTPC8SF7-XfYe0oAfEVGsxJBSG2GmZf_JxNKGSkFE,3301
|
|
224
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple.py,sha256=tCNFjltMwECvYfBzsqfjKTSAzquAaNcVVYCrsXNyts8,3734
|
|
225
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_choice.py,sha256=62t6XHt9AUUf6vggKsUckcit9ExaTBdczEmmjezG38I,3094
|
|
226
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_cpp.py,sha256=hEKkZXmOBYLAspDcBGy_upgR5Jdu7C79bc3tYO625vQ,4356
|
|
227
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_go.py,sha256=-nny_pFldms3OStnBYlAi0tyGGdSx_plQNEL3nHdc4s,4349
|
|
228
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_java.py,sha256=-N7eLd8_EnhX4ZjglRT7eL5rmLktJDVCDVBx3uaDfbk,4363
|
|
229
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_js.py,sha256=iXQcXOZtZs7f1KjWMeNdfEvCCxtWWmzDANfp24GHEmw,4357
|
|
230
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_py.py,sha256=TjowpFLoraDyppMPa7375Ol3NwGs8H1_E1xa68LAhxo,406
|
|
231
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_rs.py,sha256=7GtAWK7iig8vaCW-lI77HXOQCSnuXiN8q16XuvrIyRo,4351
|
|
232
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/non_greedy_robustness_agieval_aqua_rat.py,sha256=zcyMAZv7LDOgNdsU-gxVX5PaL8W6IYXOFeaP_l4ZqFw,3461
|
|
233
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/olympiadbench.py,sha256=moW_Y_4Tqgnr8bsFtWXmjkFmEUjjjtmngD-caiqdA_4,9644
|
|
234
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/openllm.py,sha256=OLl__FMCKuOIcAq7-nHif4uinknDFQjFKy1yUFaDvhU,3796
|
|
235
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/option_order_robustness_agieval_aqua_rat.py,sha256=GOGncOk_KiYVVjNXCgUaaVgxiowRF8avM8PWpaUng_o,3479
|
|
236
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/or_bench.py,sha256=R_tpmbnxXHAUEb47CA3s94F0LcjinO6mB_WMZrl2I6M,11101
|
|
237
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/penn_treebank.py,sha256=hp1ct35XCY4KQLtJOA36t7K5fl72o9CHDF374oJzBS8,2601
|
|
238
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/planbench.py,sha256=c0dhZOub2H9knEGC0eUjDO6dZJ8Ss1R82kT1Jms_VNs,11809
|
|
239
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/polymath.py,sha256=D69JfFd7NPoJAQ-AUbuvqvWMHwLS1TJMhoPrXqeNY3g,13058
|
|
240
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/prompt_robustness_agieval_aqua_rat.py,sha256=g-h6DrOIe4zdNfDUUAqoE90dgBUFbIiEZcxH2Aq-Z9k,3427
|
|
241
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/pythia.py,sha256=7cTgI1og4b6qSP3LCkG6m9FRnkmP1IzjT3O7ltis7II,3791
|
|
242
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/recode.py,sha256=WDeylMHDuZWoEMcLO_bubLxCJBR2X0HaXC8B9ri2jwo,4536
|
|
243
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/refusalbench.py,sha256=My4ruNBb8QeWArfkzt7cD9eglSAEiPLjPMtP-tyALUA,11940
|
|
244
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/scicode.py,sha256=RQUEzLatM8Dys-CWsMahKWB_eAILKTS8exc99REG0Qg,9495
|
|
245
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/self_consistency.py,sha256=QclAgUIwl0YPTBKHtlOyI1gSzdHeATmL5-kxgB16tPA,3035
|
|
246
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/simpleqa.py,sha256=mRsR-tBOfJXChmhSfrbX3J6XWF0P0l3iKR88mK6ov2I,5008
|
|
247
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/sorry_bench.py,sha256=orRjSL-mfjtXBWpwd18yNo-XVdUaKTK0RpWLAuiKLhw,8348
|
|
248
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/stsb.py,sha256=PB7T85YTTN1W6LvFa61ZK2OI02NZqdOsGi0AyFew2dY,2544
|
|
249
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_glue_lm_eval_v1.py,sha256=5sjnFalFsICF29k3DIyJ-PGM18yCA8IBQNsJAuZMWBc,3858
|
|
250
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_glue_lm_eval_v1_seq2seq.py,sha256=ncVz4VNXj1Be8b-oWcsC0n717j4eziW1Jj9N0vNLt5Q,3890
|
|
251
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_glue_t5_prompt.py,sha256=LqhHDSpAqDfgaWnzxmBXuxKTFcPXWwh98Jq9oGejPW8,4935
|
|
252
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_gpqa.py,sha256=dbqbbz6Yu-2pN5uk4FrHDRlpYMP9aQ7qCpw1YlOtnhQ,3233
|
|
253
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/swe_bench.py,sha256=pWTAw4T_abjQ2EwVVxUsuHTdX-N9A-yknDvMNSZCgzY,14765
|
|
254
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/swe_bench_verified.py,sha256=vWJRj-L69W7hoIs0RpBlbdBI4duakH2uH4nwIaCcIpA,5406
|
|
255
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/sycophancy_eval.py,sha256=qAkLC9LxYQIyKc-k-v1pFdHcDzfT7yLhDpX-9l6eIhU,7656
|
|
256
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/t0_eval.py,sha256=1eSUGHSbczaaAYWov4WhwgV0p2bHP-B9BOMUdhIXin0,2577
|
|
257
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tag.py,sha256=IBphPJy8Dz0FgEF0WIDMPdcErMiMnX_UbwEBF-hmffg,3572
|
|
258
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tau_bench.py,sha256=rzUmgixtJGa3WyuTt_TsJyQdiYL1nQ6aCVKhMYKzbMk,11306
|
|
259
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tmlu.py,sha256=E2_2XoupBm_LdSiX5Wi6BYabHgXabItq4r8lKh-c-Mk,4550
|
|
260
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolbench.py,sha256=40my4F9l1sIAgwXJPFWjRIBx05-CotsH6idQnDIZQ_U,14102
|
|
261
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolemu.py,sha256=91KVi6pB4P_4GR9LKeCAqOFUWiyTPtgDh4YPYlWGyzE,13678
|
|
262
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/travelplanner.py,sha256=x2_P06ZEQFT2tRNtii3NzWXCH9ItQVZG8t4GN_vRWYM,10133
|
|
263
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/truthfulqa_generation.py,sha256=TW1T3_VVfZ3RAQ73xKojUb2sI9eoYMbEEyBUKdqtFOg,4591
|
|
264
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/unfair_tos.py,sha256=UnZFX-O-nTbNn3Cic7kZ-x0msQCxC_CEMfI5Ft2OR_c,2805
|
|
265
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/vaxx_stance.py,sha256=fbZGr1NT5vjei325PqBW4h_MwHyn5b5Al1Pt_EKvcmc,2892
|
|
266
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wiceu.py,sha256=-7ckDIwPMafWk4T-D-peU4j4DbJCiZYGizE-AdQhJE8,3167
|
|
267
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wikitext103.py,sha256=Bedc7Yd9RHgqf0ZyCBfCQZM3FkSZXqbLZlxqmBEpJsk,3190
|
|
268
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wildguard.py,sha256=x_KmXVDlgDI4bTcII8DhiHyws817qimyUGvC2I_7nzI,8984
|
|
269
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt14_en_fr.py,sha256=Il8G6_AZsw7rusV3WM0LGpL0-aj1ryLs6Jf91AaWIkU,3299
|
|
270
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt14_fr_en.py,sha256=8qy9mpeQqZEsEGZXB4dKqN4Xxhph4mff48J7zZLnnnw,3299
|
|
271
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt16_de_en.py,sha256=W4kz7BeFxAIg3uFxg-og6qaNreULSWpmRgeMu8bt8HQ,3193
|
|
272
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt16_en_de.py,sha256=Z3VhStSCCQk1aJnrN2NeZikGZe0KbV6DjDr3tzQM-lo,3192
|
|
273
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt16_en_ro.py,sha256=dlvYhgjZyZCnCb1FV0LQzPbswZbKVxk8UQfDRtradP8,3200
|
|
274
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt16_ro_en.py,sha256=hEQb6pcfKUmdulDBStxMYDfQdEykb_YT_Pef08y1DXA,3199
|
|
275
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt_ro_en_t5_prompt.py,sha256=SIHsGosDvHnK_4C3XRHUAhrRN7wFFNAKB0PZ31VPfc0,3241
|
|
276
|
+
wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/xsum.py,sha256=zzIp0MeqI7XccHlNEaqL0w3OR3Ukt6lqMYVZQmksAMs,2907
|
|
277
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
278
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/atoms.py,sha256=u8RTCYlpUX7X3DHLx9aPlid3xrYPQK6f7Kymj0ps7Ug,9056
|
|
279
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_manifest.py,sha256=a5Tk36Wm0a7bA0mwtoqnAGP1muYUoGEdW52ZhPMyrx0,29947
|
|
280
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_registry.py,sha256=Wq5_AOWRUO2m387YaLmzWPWfMmwT96FmaAz0N0_JV6A,5255
|
|
281
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_pairs_generation.py,sha256=x5dWdceSnMetwmxliUFmErAi0DHXknSBUQ3SSXBznh4,2090
|
|
282
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/__init__.py,sha256=yxkFjHDEpD33FH1-shObI-jsperBSADjTqZxQrzC3QA,16627
|
|
283
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/aclue.py,sha256=jreFoIxdiaOZM-qsv_L8dAkJxXm21ajQR0XUY8LJP-k,1321
|
|
284
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/acp.py,sha256=2txurLRPPG4xBLMd0a8LJjF4XAyQHmQon_gp4PbTl4g,1902
|
|
285
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/acpbench.py,sha256=8k0y8xwwEAsZ8oIYKvDpNgBCgSoTDQ9IDLLtHcW7p-U,2423
|
|
286
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/advanced_ai_risk.py,sha256=cpvm-S2UITNjxB2j0-cP3lE8nXQnL63b-adqscH-2OM,5075
|
|
287
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/aexams.py,sha256=igjDewyTuCq4PenzkEFHW_Xqm8StRvzfGlj8iZQ9CP0,541
|
|
288
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrimgsm.py,sha256=EeUJKokyNZNDqJB3ce1ebm1-MXXMX6HBCQX0AxrqI9s,306
|
|
289
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrimmlu.py,sha256=gBNawvo0MvrKU2lr089e-CMBRqzkprr_WXqgb9QUo0w,306
|
|
290
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrixnli.py,sha256=TYGAY_Fx9MwWn4Qy8xL9PE3ln0ty-wb7Inhtm3SCUWk,235
|
|
291
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench.py,sha256=IHPw3Cpv5q2DJ1Snh3AVmayBPNc-HFj_Jhcd6k6uJ-E,438
|
|
292
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_adr.py,sha256=TubmxD9KiQuuExJCnOnqcXqO1WAyV5YIKJfbnbrNuKI,252
|
|
293
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_afriqa.py,sha256=aul9JhLJq1OKl2JlwNaPuQZg6MY7WsvXS6N-O7w8gOw,261
|
|
294
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_afrisenti.py,sha256=knwROtJyfbVsGbCRH8zh50oUTygw5GkAXORN5iN1Sfk,289
|
|
295
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_belebele.py,sha256=MVwUau0u-56cUOwlmIuujeAV5S4gqLlvq-kICwr6YIQ,284
|
|
296
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_flores.py,sha256=Zsc_k_qH_HdG54pOOTgtR_n25E51fnpOrxkzDGq1NVI,274
|
|
297
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_injongointent.py,sha256=cQW-WexgKjEOjStmNTegX9rOTywBn3IJJvNIb4zlWUo,309
|
|
298
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_mafand.py,sha256=-srB0z0rfK3D_d2dFIeqkWqcerQsNSXNHqE_2MlxA1Y,274
|
|
299
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_masakhaner.py,sha256=1iYruOX9fZUnWQtrjMFh2YX7e5FFtcB_RFcSpsJzuDQ,294
|
|
300
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_masakhanews.py,sha256=rfqGxVIzM6aXg2MQFE-ZWXx3F1eHpHTjOQqHbBel13g,299
|
|
301
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_masakhapos.py,sha256=Kiw1_NhBfroC_g_vxksygxK381Cx6FyCqByPYIsZZd4,294
|
|
302
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_naijarc.py,sha256=FyK2UA399lvSzNKO3bLh6vazRKjN_nER8JshHX9pj3c,279
|
|
303
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_nollysenti.py,sha256=WoAX1qjTh64nRXYDXWKnwVgj12nm3qEL6Bn3fN3INZM,294
|
|
304
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_ntrex.py,sha256=nIkFBvD5YqO2ew0-d0ns4y8DwW8mhc5TDesknui7PzU,269
|
|
305
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_openai_mmlu.py,sha256=GvsndFbxPY5hmPjKVUBPnbxRUfw2uhnzAy0UJ78223Y,298
|
|
306
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_salt.py,sha256=jhW2mYhGISokNMC-eS1TmlbQdpUmpeR2WqO0ByRo44Q,264
|
|
307
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_sib.py,sha256=_ZDBceyVOGn-urUsEZcsmPJcy4sHaZtb9zbmVhykV8M,259
|
|
308
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_uhura_arc_easy.py,sha256=64C9nvJsu6GVMp0ua5nMWZPylt_3xnz6LHf2MEqM_8E,312
|
|
309
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/afrobench_xlsum.py,sha256=TFYTh8mysh_Tko_u-2bp7z7xenHsRrfYnUf_kRrw2CY,269
|
|
310
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/agieval.py,sha256=TxW-eUZpLm-NE9pnHSNZ2KoE0Yh6cNDnXw5i6Q4iVac,1866
|
|
311
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/anli.py,sha256=6nb5vacMo4dXia_sLeLpwj9lhWLu2AyQsQcCI0hRK7Y,215
|
|
312
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arab_culture.py,sha256=rNBkeqYZpEHS2km9Na2EQ61iWCAMnV_fZky8vIqhDX4,1387
|
|
313
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabic_leaderboard_acva.py,sha256=0jo1F7XoquyYdzDV4iC69jD1YPSM4EJrxMbwZjRG7do,7581
|
|
314
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabic_leaderboard_acva_light.py,sha256=j-D_DICXClTULp5BNbyI7VltD-Ix6wGs4S_eXNjDt1s,7593
|
|
315
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabic_leaderboard_complete.py,sha256=rFhk-DG5m_0Z27fjT53zKrMqcNgrR5NbwuGokR7WpaE,2132
|
|
316
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabic_leaderboard_light.py,sha256=tZdvutqxZVmZxMqdVZxvP93NgU2PFR4M0iwd9CyhEds,9282
|
|
317
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arabicmmlu.py,sha256=0b--BXMHjh3hjcLPWB2Y950sH9CH05qZcDTG0lQtTQI,4661
|
|
318
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/aradice.py,sha256=vuY5LcdNXtpKebQvnblvoj_lDlYOZiLGnm2knTW6pX0,2168
|
|
319
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arc.py,sha256=7YOQGtpn546NDKCb1unMj5EHYsMYLDXjU3sFhPZ2Wos,4559
|
|
320
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/arithmetic.py,sha256=zbfwvNhXawPxEdYQ9jQ-MnEwI1ZNIsS7RxztkJYgXmM,945
|
|
321
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/basque_bench.py,sha256=k_QsdtmGDw3Vk8Z3smZG_gagJ-LSJMJGARcNN1M4fN4,2577
|
|
322
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/bbh.py,sha256=sdTO2A1cDgyIORBiHEk7WlGK__h9gJsccUCrDQz2x94,8786
|
|
323
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/bbq.py,sha256=BiE0lCXtqLqr4086ZwvUdE-RDcj_WjOE9-1RNY0mH0A,210
|
|
324
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/belebele.py,sha256=U8dMjWVatzBiRB5fhjllWxjwAl6XVG7zm_lsTUieFVE,20448
|
|
325
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/bertaqa.py,sha256=bx2dbYc3HFb__oJ0p1hkqtm2bKzcA6jiOVlGIfmuPGE,1382
|
|
326
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/bigbench.py,sha256=wTVFn18R2Z88umFsoaUCUQYpd0ezJf6P0RCqMJFaem4,27551
|
|
327
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/blimp.py,sha256=kTt8oD29-ryYriuNA7b2rNK6ZdNKE9FWRhkVTz1cC4A,5466
|
|
328
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/careqa.py,sha256=I_ZfnRbx1a3wg9jK4-QNQE8fP1qfqz_5MrrY5qx2Txc,225
|
|
329
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/catalan_bench.py,sha256=TPUzqtL1du59Dt9EoHaKriyoFromWMAoAswjy52lOeM,3167
|
|
330
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/ceval_valid.py,sha256=UNjahQmaNAA4Tbk1VRrwjwskbh62xZjNTcCh75Agu_c,4229
|
|
331
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/cmmlu.py,sha256=HAWYWsdvK2s47rmzwf3c1XaOs-NwkVM0ovWByD6klSQ,4795
|
|
332
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/code_x_glue.py,sha256=nAY6LGRl2Cegbq6_DDCFtI7LsnZraabCzHdhicOIkpQ,710
|
|
333
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/copal_id.py,sha256=he7n0aGC0XjMR4qa2gm0kuovyz0_x1JLEYJNPfXDmuA,372
|
|
334
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/crows_pairs.py,sha256=7CTdKKc494cmFzBjofRyp2_dlNwiyqBPiuwlkDyXAjY,2136
|
|
335
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/csatqa.py,sha256=Eo22i1KWxQX-PNZ3nGPm3UQSKmP8oxrY3V8zchNwOrc,571
|
|
336
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/darija.py,sha256=rFyqUWJWCxdDKbiuCNgFccam0idwTqrzdlT8E1k0MLA,1894
|
|
337
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/darijammlu.py,sha256=jLUs7QBunFZmitxCN9N2MtxKRTWWCOpA7_7nI9HTM7U,4139
|
|
338
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/egymmlu.py,sha256=TU8dB_QIg8V92CQ3D_e81pqgWGYctulyvmd4tq7J-tI,4054
|
|
339
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/eus.py,sha256=ixTudUA33UTuUKDKPh5Cjwctmg4uNV3T300yPv4Ef1Q,5510
|
|
340
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/evalita_mp.py,sha256=_LdM__wODPNdavTHGgLqF3dA_vLHC-ewmydHm8Bs7q8,6644
|
|
341
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/fld.py,sha256=FLzUNm80tvs-htbAfm4mNq-HSiu8oEYufTzstf9Nk4o,210
|
|
342
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/flores.py,sha256=A75TUNok5NVIzpCO3VwI9m9Izh8pXtIeNdr1v38Kjec,35617
|
|
343
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/freebase.py,sha256=SZoUpwil8030dBz37W4D46Oz0NUjmS2XyAWhKNhqKN4,235
|
|
344
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/french_bench.py,sha256=Rqlk55ooPSoKOp1vQCT-jPsDqinDAC6ot5YkCIy5XwA,1554
|
|
345
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/galician_bench.py,sha256=csNGQJO-ML0ViEuvWGDaa-V22tdOylnx0cPsSbbTdXU,3073
|
|
346
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/glianorex.py,sha256=diTVUzsCmpkEl-xfnMCmrUChExkEqRPtQpZ2qHmN-Q0,372
|
|
347
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/global_mmlu.py,sha256=hnv9ppNUeez1SizAYPWvaVz6XShHGp-f1lhDkNjqp8Q,8628
|
|
348
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/gpqa.py,sha256=IGJIxS9Les-J04Zt557ni25GAokJgVoIhqE_NRgq1vE,1416
|
|
349
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/gsm8k.py,sha256=ZhVoA23IztCsMv3l35L87VBOFm2MoSIZfThWWxNXVg8,220
|
|
350
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/gsm8k_platinum.py,sha256=r9mhMUWUmgbgVK2m46NhQ_QMign-L4or-f0rfFibIvc,264
|
|
351
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/haerae.py,sha256=I71MTrDBW9pk2co7wRY90CtIs2f4JTi-DdRyOz4M69o,563
|
|
352
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/headqa.py,sha256=96jSwl76X7wGSdS8yEP-9X1V7sUJZmwqjsFvZe_svHk,339
|
|
353
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/hellaswag.py,sha256=AZvyCHCrWbq6Ms6qds0kadqvbMTHNFLX3yq_Rv_nfKg,1980
|
|
354
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/hendrycks_ethics.py,sha256=KNCrcp5CzVH4xUTaAEYm6WMbG-T8axw20uYqIwkhsWI,683
|
|
355
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/hendrycks_math.py,sha256=lAVQ-mRhhoy7xqRPPFdHFizE5-xzhNs8nOSfcdjd8s4,264
|
|
356
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/hrm8k.py,sha256=8OMa-PO4Np4iRfMhntf1g148Xgd0OTToShoOH812hjw,861
|
|
357
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/inverse.py,sha256=JTWvuZAXZBTwnGVcq4BQMj-46N-5blx8K_NllGy9eQo,1456
|
|
358
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/japanese_leaderboard.py,sha256=oU3bkWfkUgnl2uLD-GqsK06HYofr1rtY8_mYqsTOUuQ,1376
|
|
359
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/jsonschema_bench.py,sha256=vSZbZiDDQ-LvbvZ4vanb3stPS7o6bTeZseUQ8Ju3VCk,274
|
|
360
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/kbl.py,sha256=F64FV-_4Wk6dtt2ZdEFgDIyETsM3Vw9xfwMdjaVET7U,5272
|
|
361
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/kmmlu.py,sha256=Y58gLEx9Q0ZNddE4xHMxqeHig_Hx-tNtsiW9VckChFY,24414
|
|
362
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/kobest.py,sha256=o99MuIsjs1I52ERdqsA4q3NxvzZJIn30Z2YhFaj5koA,529
|
|
363
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/kormedmcqa.py,sha256=hvJj9iU_7ENv7Tx1AGeQwXxbUy7wTLCL50i5afwlPn0,245
|
|
364
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/lambada.py,sha256=TNh6u7Q8M7AWonisRUoueuAEdMy2ztfIf0dAaaXNDDA,2146
|
|
365
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/leaderboard.py,sha256=QDcVieNMmPogUpQj0EMTTcd2ttrZw40LLS4vOEZneTM,4153
|
|
366
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/libra.py,sha256=TTXTuQ3h7giyXwE93Wbc6nO9suotWmPnxOKni7Jkq64,220
|
|
367
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/lingoly.py,sha256=SeSVwpPcCoFPoOelrCl9XiD1it5otj-WWUg71N8g2bs,362
|
|
368
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/longbench.py,sha256=3X2mxA0X8O__ObbVohoHbrLCD9SDvpgBg_F85UWUqNM,240
|
|
369
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/m.py,sha256=u6z5soy4JV-8TFYaJdxFF8wwfrNx6G7ck21NQ3sSbqE,2084
|
|
370
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mastermind.py,sha256=rSiBDFFy-to59CZ7eLtR2r-IbpkykJDSGvbvcqzs4Lg,245
|
|
371
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mathqa.py,sha256=qCK-GrMTjzTs3sgxPwBKDAmU_vuNQUS8v5qZ0gXdkIg,225
|
|
372
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/med.py,sha256=BU3WfTzLQCsAyGKKCQeKxD8a4sojcDTsOqbXnqZfYS8,1628
|
|
373
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/meddialog.py,sha256=C_JHvittLnB9AHA0Z9OhVCwM6kSbCYXzp5DwX7IA67M,481
|
|
374
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/medqa.py,sha256=GKZyG1RPco_aldysYhhDBmHnFrxEOfCHSfb-MQ_hBtE,229
|
|
375
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mela.py,sha256=VYRpUqY9y2kGESnqIe4qlM4h1mEnO6r8N078n3IhtYU,674
|
|
376
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/metabench.py,sha256=XMZmtzU9820jqDeZ9jY_HrdRw0MiR0Rv3vN_q_Mf53s,2282
|
|
377
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mgsm.py,sha256=3J0HoSvj-zqVPSseEtyuYt1iKQHEkODKlOV71vWIenU,2331
|
|
378
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/minerva_math.py,sha256=MuFC1fZoU918mOa9WIHGlP23Kw5NaSQyHlFC-4oZixE,837
|
|
379
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mlqa.py,sha256=glrPe48NOl5AKIsg5ZC7UnKDYRfCrsXlmkpQGOE7Bb8,2861
|
|
380
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlu.py,sha256=nK8GgXuSb8A5L2yQ8OFiieBvtRQO2ZIQkT8Q7r_nZ1Q,5161
|
|
381
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlu_pro.py,sha256=xUueql9tBJWYmNg52Sl0seUuFmMBpp5TY2Mk5mXRCxU,1186
|
|
382
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlu_pro_plus.py,sha256=T6cx7TBuH66EPddZJFC9kS5_uvQv2OpgtIDibGqnfxA,1271
|
|
383
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlu_prox.py,sha256=1ygyPFiqhtz93Q3iMaiZTJcoqjT1s5kNcpXFB02bwfU,13341
|
|
384
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmlusr.py,sha256=AWz8v7FdOzmdsRkTYkEfGztVVTt5GjbBI950UqpFZvw,225
|
|
385
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/mmmu.py,sha256=99ApR5ukDUKDU3EE88RhrsUOwytA8--Mx68LQCbIJNw,2616
|
|
386
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/model_written_evals.py,sha256=K05a8SFb-f-eF1d6PEQidbe-X84dp6ecikhkRW6sznE,288
|
|
387
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/multiblimp.py,sha256=EWOyj45fS-ZJiuGA2pJHyNArpUBSJ42HHofh4UUmqAg,7360
|
|
388
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/non.py,sha256=cjMTZBJFHHCrF-T9AzctOpIZ01cqTzhZtZAKXoRN2n0,1384
|
|
389
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/noreval.py,sha256=pg422jYIutSs_xyfJYvZNmC5MTJ9UNsYrheKJxAzHos,10342
|
|
390
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/noridiom.py,sha256=X2XFnPbNpMCM_PI7vAllIUhHms0IL1knDOABDl_PUKo,1117
|
|
391
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/nortruthfulqa.py,sha256=shUWL9Fz56DJ94vbXaygvxm3t23znDrGjRUJZM0H_As,2309
|
|
392
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/nrk.py,sha256=Kx-sYWuQM5H5KdXGzG7SMrOcm7pjlx9FnjKPgPZK3dk,1179
|
|
393
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi.py,sha256=kxD_mmW4tKLY4Wx3U5JzemTNt9z6t-UqfXi762EEdyY,220
|
|
394
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi_arc_multilingual.py,sha256=NI6jZ01pY5Z3ZyO7Y9L_I2HLjKNQjjYRQQ5VfOlQcgw,403
|
|
395
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi_hellaswag_multilingual.py,sha256=WhujIN7w7sr9YY5sOemi-gkGc_FS7dMwiPBtrFVweMw,1198
|
|
396
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi_mmlu_multilingual.py,sha256=6-awa8sPH2Lzw4hHrAcg44TCHz0lgtU2wWWiN7oJRzI,1090
|
|
397
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/okapi_truthfulqa_multilingual.py,sha256=cOdi3vh4wxIv4ej-7O-K7fk1e-W0_2-tpHCAFjg42cY,2107
|
|
398
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/paloma.py,sha256=HsC86ThIehSit5L6WmeJ3WxzYFsNC5GOxVw2pgkcuHc,1325
|
|
399
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/pawsx.py,sha256=QtgMM81o8y5b-rOc5EoQIxSsfgsrg-yn4fiygobxNZA,220
|
|
400
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/persona.py,sha256=e6qIse5XZHvcGFa3gc_vB6_e2AADjtbwQvh7vbkziAI,12876
|
|
401
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/pile.py,sha256=gQIzP9gMvxDZXMYbZq-I2SdFLKlOF3h3lpx6C_dWy8Y,1511
|
|
402
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/polemo2.py,sha256=EPGAkq4z39U1o7ut21Yl_fqi7KueHMH2G4Kc7jfxLB4,230
|
|
403
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/portuguese_bench.py,sha256=N5_ttKDHVEXlVjltpmJfJgexOfDXPRJ3Hh9X543jmwg,2236
|
|
404
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/prompt.py,sha256=bpyFJO1YrFiwO2f1DZmh-diiUbOB8UjrGT2oH2TbQJM,1343
|
|
405
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/qa4mre.py,sha256=tF4SqQ5lteeLwx5m5ir7hH_m5pZJAyR5e1UKxXsaSVM,402
|
|
406
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/qasper.py,sha256=0yjWxqTC5P6VQ1rWzORFdONMLnZFt6FWb1__oo0rsPM,347
|
|
407
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/ru.py,sha256=HLsjyikFgilRkVeQm5d0fMjvYQA-9wBWWQCcUaG32VY,835
|
|
408
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/ruler.py,sha256=uLm6N46UFsMzM1XWWNZjFqON1bOyhLzv2auDIYGnfqE,220
|
|
409
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/score.py,sha256=c6vo4QUri_qAdE_XEQyR3qaXfJRTyGzpiK-lETEDkio,1076
|
|
410
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/scrolls.py,sha256=FCJUCO-DPbI8MsAIPpjMcAtvWIq8Tt5lBuG1l49E-Rc,230
|
|
411
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/self_consistency.py,sha256=BzKb8ZFeYlbgTU9_GWwLpkFOVX4EG2FA7jk31oGYzYY,406
|
|
412
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/spanish_bench.py,sha256=XeQMJTLWUrOrBB24nshRHb8GM25xC6oQgGOk3LkKy8w,2744
|
|
413
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/storycloze.py,sha256=m6iuzSlVyn1JN0yN3QOHvkcA17-njh890RwOdnqpKFo,247
|
|
414
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/super_glue_t5_prompt.py,sha256=TRJv9l-zRi-9bs7oQMhDBLUibSbVshihVPKuQfREpLA,1077
|
|
415
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/tinyBenchmarks.py,sha256=FaU5zkZjJCNxP6b-j2UIVHUtggbKUZTlTkw9gFRqrm0,265
|
|
416
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/tmlu.py,sha256=D2H6RHQFVWhFZ2Z8pWHMEqW10TxlDI3NMdwfca_5Oac,215
|
|
417
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/tmmluplus.py,sha256=wsvggPktE0jkSan2KyLco34qR7yze4Fgen3HRMPNm0Q,5955
|
|
418
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/translation.py,sha256=HWyJdTJL-ecVu_pa6HKdVhIo92oePCDBxbxxcJJkNAo,250
|
|
419
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/truthfulqa.py,sha256=nxUJH-jpElqSU6jkkg-YgHMGOIDqw5RBj9v_pV7wz9w,7424
|
|
420
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/truthfulqa_multi.py,sha256=P9z59LiScSmpxrpm5Ay7JwPJIY4w_MNouZpShYpH3eo,1624
|
|
421
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/turkishmmlu.py,sha256=tuVJu8x5I_e-UJgC8jP2h5q4Nye1jdmjzMEUyJkxwAM,2176
|
|
422
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/unitxt.py,sha256=ZCnls8Pbqrte7CmARZOQrVWsf-LpdChLPOgDQnR82X0,1078
|
|
423
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/unscramble.py,sha256=J8hNpbDcvgJA5xTiDnRU7QjH-6E_k2KoeUIlfdY8oOs,245
|
|
424
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/winogender.py,sha256=m0qTd8AyKj36WN3TtB6iKDuTx4tzWWKB8xIihNt4k78,764
|
|
425
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/wmdp.py,sha256=qsiUynLU2qNdfNL8q2gIpCBY_xvQOVD9q4HRmkU8eM4,374
|
|
426
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/wmt14.py,sha256=yEr9PGLc00t202EocD_g8lmrB-ZCg9Y-krg-B52daSQ,615
|
|
427
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/wmt16.py,sha256=fPGA_y5Awe1CPIGUCXNW8MbrZa8z54Vwsbu4vqAnaf8,957
|
|
428
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/wsc273.py,sha256=OUQZSI3gN150QLkw7lEilO6_Vxql27T4HzVg7kKcZds,225
|
|
429
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xcopa.py,sha256=qOdg6OfwebOax5yGbvcP-GsTv7wa9H-Ek5QiLeVhqIk,868
|
|
430
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xnli.py,sha256=Z13nYRZ2fUv5c4oWRDDf4Pb-IMTM235wl7SmMRRJPoY,1198
|
|
431
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xnli_eu.py,sha256=ErEgwkMgLiV_qbgWICeC0zgS8mi9Ln0atCuDTafQlPQ,397
|
|
432
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xquad.py,sha256=RfT-MtFb-CpAyF3svRJ10gcqqfNxxKYJhd5PKpLwuVg,922
|
|
433
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xstorycloze.py,sha256=VLiKWLfyDPJnk_L_aLchZVV82KCIU9T6IgSvzYS_ZBA,1186
|
|
434
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xwinograd.py,sha256=RAkchOuJtal6YXe7U1kiNtLJOzepczRuRNqzv3CNKm8,636
|
|
435
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/__init__.py,sha256=qjtVqnPSAZjOg-ME9QjeOEkgJR4rImyfGImxEwAEfmQ,7278
|
|
436
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aclue.py,sha256=uGbZtgDttvJE5BeAqpkUWOEoUVVDo7APzFXXjAh-HQg,6783
|
|
437
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench.py,sha256=gX67-4ZALgLegw7rcKSuWfExERxPqiC-1ZAX2rahjaM,8489
|
|
438
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench_hard.py,sha256=SyIMnyCm5BYWVZmA0x2KzO8otv05e5FQ-vq9mM7u6ec,7582
|
|
439
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/advanced.py,sha256=KFexdT3Cmu-cVGzYruT_vHIHJq4-cdhDgii4UEN_cPI,5298
|
|
440
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aexams.py,sha256=3xIlyGLP9TKhBo3Z0OmLY1a7bQsshnHgXFTxkt1RKrc,7275
|
|
441
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrimgsm.py,sha256=ozlEmQHqxsaiH5jb1D1WI9aToAk6psQkduDwq737F5c,3527
|
|
442
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrimmlu.py,sha256=-hH6AKLfE7k09FP4Zt_FFeapGp52BSpYaduyFVhDzKk,4139
|
|
443
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrixnli.py,sha256=aYpA28yQtnwMfstzyZJDakPhV18x84fVJDIKMZ_d0aw,4466
|
|
444
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrobench_cot.py,sha256=XXg_BHyIHOt_NezgVJ5K8aplTR_rTiX1Eg154L3sTI0,3319
|
|
445
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrobench_mc.py,sha256=1aRM-_bxDOqWWc5WCfaes_WIbgfs-zurYQkP_PxR660,4114
|
|
446
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ag.py,sha256=4lDiGXPZHhThl6fa6DKho14cGEWLnKDPP_5fdTO8gUg,5438
|
|
447
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/agieval.py,sha256=mvO8MHp-Mi6w4pHE0yd-tF6UFfcj7shWUSdV1QsH_y0,6018
|
|
448
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ai2_arc.py,sha256=0Rrn2efrqGl949i30Y83R450uoc_y8o4H31qj46dr3g,3878
|
|
449
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anagrams1.py,sha256=VMEhdDEjXyQscQGSXz98H42LZ75XTuS-QcrARzWx-QQ,3141
|
|
450
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anagrams2.py,sha256=gnfUcTl26pUmoYPGDg5SUBasrmCK1ggB_kIQYsijNeU,3141
|
|
451
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anli.py,sha256=YunlhuOmXsEO0ic8AzTgOz5I7LaYib1IRdTwFTllErg,4808
|
|
452
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabculture.py,sha256=_tSYyWpXsmFOMe1dR4DLY9cG49la-QC_2J0s7XG572k,7337
|
|
453
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic.py,sha256=8UEvnJ_q1orPqcxOlKCmbwVKrXK33fmimLwa2YvDsGM,3766
|
|
454
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_exams.py,sha256=Kxe471_ial8PtoJD607yNBxijquSVQMv24DMFdAt3T4,3953
|
|
455
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_complete.py,sha256=1HbKhGtw6YL42WRTjGHZbLxUMj5Hp26gFrHM-2Vhmzs,6893
|
|
456
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_light.py,sha256=cicxXygeMxPteGOECVtraOzvdZRrbd1YW6lw4stvXAA,6863
|
|
457
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabicmmlu.py,sha256=9JM3zPOImCsegFxxVLBkd-QwH4JylErdDTZZ_vOx2s4,6754
|
|
458
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aradice.py,sha256=m7VYmrliIxYsRSnmk7ySVaMtx8sS10esNAeeGC6BYSs,12083
|
|
459
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc.py,sha256=cJTEsh7gR6xXz5of_Al_sVxduGm--tXAkjBS_VRxZng,5238
|
|
460
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_challenge.py,sha256=TWSj3EjOBZy6BPxFDwxdyplzOQCwtjspZHuIEbG6xng,4121
|
|
461
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_easy.py,sha256=uJc-wlODwrMcrD6Edn1ZJMIJaZfNMA-thPe6jOMFnWo,4076
|
|
462
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_gen.py,sha256=Eokj_PknnpqUK264I68i0DiAjQVcCQxH5lhDPIZ_zes,3881
|
|
463
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_mc.py,sha256=TXvAaLY4wAsHuozHBJ4Dp7ryOG9bA4B_AyhYzS-8WRY,4121
|
|
464
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/argument.py,sha256=GJt9XdLoXcnJ7_Obat9nK-CenPZNvIaFva7e5XeeXf4,5494
|
|
465
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arithmetic.py,sha256=rz6kaUDq-4HgW8ERh48fFWoHfBQZ-jClYD5YB1PIZa4,3880
|
|
466
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/asdiv.py,sha256=8xtQLBuSM31XXhhDH_PvsWH-D4pQrnY9Q7plGafjdF8,4215
|
|
467
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/assin.py,sha256=w0CEpwyxBfHa5SDXEUc-rTbogw9f6vqT7gPiND-XNa0,4158
|
|
468
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/babi.py,sha256=RkooITV_aXF58uK69v3Ohj15bPOJS5yYTzuDtLZfNSE,4020
|
|
469
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench.py,sha256=zZEJMU0SroZMWxI3dkFrlk5CdpZeRJqLbzwxDdGUWQY,6065
|
|
470
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench_gen.py,sha256=Er8ufToEqS5vRAbl5eHTz7Ihwrv7WYCHdm_WN-GOqWI,5983
|
|
471
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench_mc.py,sha256=F_dnN_Yips5nEX9ZW-nOM4HvA5a6XQ4-WCkeUsig_Ac,5383
|
|
472
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bbh.py,sha256=QyL31K5p1pSAoOeaThMXcqjXvn5ARZeyn458y6wIpVw,4730
|
|
473
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bbq.py,sha256=VRMCL7kVxeQidjiQpkMvnJeq6sUyrzGrerLjaLcHi88,6730
|
|
474
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/belebele.py,sha256=ksyM3kwlAXoATQxNel6FFSLMyDvxX2uxmevUDUzBCBM,7253
|
|
475
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/benchmarks.py,sha256=uNsdQoa3vnF7A4E-tixhO9RpCiYv0t5BCroo4wgHMRg,6052
|
|
476
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bertaqa.py,sha256=qX248uM7Urn88jpwh1ALnAXRxjV3QDGDnGzb-sLL854,6561
|
|
477
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhs.py,sha256=41eIftkv4auWeZjROUfu377zPcd1Q4_yBRNdDcau6Lk,5982
|
|
478
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhtc.py,sha256=NX5M5V7X0DBBSlEerITZXlc_Za3F7uOc1VwtDuX-TwM,5816
|
|
479
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bigbench.py,sha256=Yww82fQuUbEANLotw7vP7m_EDoA9CNRWPvgrB6B-wvM,6613
|
|
480
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp.py,sha256=t0qiCdHfIxyHN7gaKd1_-Jvctp7yEHhzR_AkBmphP70,6754
|
|
481
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp_nl.py,sha256=rGQGExRVytYnk8yrOI0m7IMg_X4LPb4lbDfF3X4XMYg,5963
|
|
482
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/boolq.py,sha256=rvwWqSRlFv-Lhc7QgVghe_iEpVtNx-2WbfysUF5TZwI,4020
|
|
483
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/boolq_seq2seq.py,sha256=-B7CD59m5m7WmL6ed3Ior9Zt57kl4jTOlbOkv9p_iFY,3998
|
|
484
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/c4.py,sha256=xjbcuPLg01i9SA5W21zusuhEUfnIaI3tMuCD4dlzi9M,5786
|
|
485
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cabbq.py,sha256=uFUWHDqytKtuwbs-iFB46CCvCsniNxq9j9j-Ae2nl-s,5938
|
|
486
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cabreu.py,sha256=_dS87XtZBWSvjTyzvBRjaK-dN0Fv6kDihux5PEj7neQ,5046
|
|
487
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/careqa.py,sha256=rKnAsp5IBez78ak85i2iQo1LliOMaweqB3GIyaBA9gg,6775
|
|
488
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench.py,sha256=EycMpZ-XGkRro3DO1xUZ3vFW0_jzDtMhcCvgALlBidU,6075
|
|
489
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench_gen.py,sha256=vYrbmDvEWetRTxFYa00e-EPmcG698rWSNTj6abBxUs4,3976
|
|
490
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench_mc.py,sha256=U27A3A9nSYqKmqSPWeVBU66eIDiw4Y7UjA0YttcELEc,4150
|
|
491
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalanqa.py,sha256=q6Hr2jQCkcVUuOjLyIzBYsUXSWQZL34352jXLcXMzvk,7256
|
|
492
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catcola.py,sha256=1Qs6oyLrRYMDM_NHkCU80si6TddhiY6bsE5m2yWssAA,5509
|
|
493
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cb.py,sha256=CLBTHMpQ3RyY0gn-_rrhCgMWPNHna4bnHz6ldbkYkFU,4080
|
|
494
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval.py,sha256=8lcZUqajLrtuC_yDtHPmRgdwhH9A0YfoDQJrOAVQXA0,8640
|
|
495
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval_valid.py,sha256=euDGphqZuKuzsIcTw8cUw21NyMq8iZg8Ll208r0XiDI,6322
|
|
496
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chain.py,sha256=pTCwNeM77UhiyrqN19d3NZLFImL3oZ9-fDs9Ael9hHM,4347
|
|
497
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chartqa.py,sha256=BfKHnGYUSQuFeOg1NJK-2dmSfkV84FAE40Kj4UA7gE8,9313
|
|
498
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/claim.py,sha256=QenIkB_V5fYn_gACF0nNSWJu_wBumGPlQfhVzBs3dGk,6230
|
|
499
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/click.py,sha256=xvl2xuzFQMxa7gRrbh-pKfqAPF22LTqqZZQgQmv1asU,5938
|
|
500
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cmmlu.py,sha256=jvoB7X6gi0-tByNHlIMuIaLvQdLMW3WrsNATVuNWi8w,6661
|
|
501
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cnn.py,sha256=o3CkInmKzwZZDNFtNtRotFQaTnQrCsUjn19A1I8EWhE,5838
|
|
502
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cocoteros.py,sha256=M-MI6r1JkFFSTcQLg_wawf_QZEgODgNvbfroxr0T6PQ,6034
|
|
503
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code2text.py,sha256=4BrCmZw4cuaRFQ0gFs5pTbYuwCP7nurs-P3F6ZJTung,6375
|
|
504
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code_x_glue.py,sha256=warH1kPxAB2Ha5weOzl4dI5vGeCjYY9Kim8RIlsDHNA,4417
|
|
505
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/codexglue.py,sha256=qeZu8Lzj_aXbWBQYTtuhEwor2ZO9n7V_WelVLLX1zIw,4288
|
|
506
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coedit.py,sha256=jHan7PwlX3wyQjsW49qzAw-P7DlJPYjZQ-tEoNiU-l8,5875
|
|
507
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cola.py,sha256=dRnZPR63cbrzVyiklwJX05cMB2fUTujZhNpQC6IVG6c,2996
|
|
508
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense.py,sha256=QLNyRHimX8aVUiNuT0K--I5-vYo-FNFTnvTfGjNNGJM,4296
|
|
509
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense_qa.py,sha256=FuiwbSPi5-iy48QHLmu01DCBnBziVZWVYCqJCO8NkG0,4565
|
|
510
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copa.py,sha256=a7pe0QRisgu0ItvdAyKXZujks4Y-aofu-CnFEISKsNY,4141
|
|
511
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copal_id.py,sha256=yyzr-c7prHDTVmcErfGdYiPnjVoKfrWm6rJUXCk87pc,6825
|
|
512
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coqa.py,sha256=YR5hcHzOWZXiEjm5LjFlC0mZVAXfnk_NaZMKZgpoxE0,5790
|
|
513
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coqcat.py,sha256=ebeI1P36JuauSiy2cuk8-BZahZfc8ylxeumXkh1me0c,4241
|
|
514
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/crows_pairs.py,sha256=IMDNwLFfzLxzFlF3TIdINstY4-XIrUfrMtUNmDn7g8w,6042
|
|
515
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/csatqa.py,sha256=2vs9t9MQt7W7qfWidHfXdX51jxr_P0VH_2D1hpXvupY,5947
|
|
516
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cycle.py,sha256=Yad__yL8WcQu1BddUd0uO0D1SQGrcOjFCCd_UZ6g3aM,4272
|
|
517
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cycle_letters.py,sha256=MUH1UEnt0ox2FCkDWrfatN1uoxNm43ut1SaToBYmJRg,3166
|
|
518
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darija_bench.py,sha256=F-uJqfXHTKQuwuMidqf8GeFDRXzOIjamfKUKQanoKgE,9142
|
|
519
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijahellaswag.py,sha256=00z6_kfvpgTFuML1NGwdsNJ_GPwueuZ0VIS1ODcpdUM,6887
|
|
520
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijammlu.py,sha256=qtrgc7E8T1nzPAqQWbtpcpDM6Aa3yZSXI3VCEx5S0q8,5983
|
|
521
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/dbpedia.py,sha256=nQrrwGH9dX0HnI40czs5wa3_kLF7SX_clObmkTeqsls,6123
|
|
522
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/discrim_eval.py,sha256=T9fdZq-3lng9PhKl7C3B8MW0WVEXcz6tNT9GD-i3bhc,5999
|
|
523
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/doc.py,sha256=fZCZ4H4pob7hIAr0QIg2H_Pgz--TyQbVFYKuR8zxpoE,4264
|
|
524
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/drop.py,sha256=oQscohjTgBxuXA3oavOz_-VdC71YYUD3pNOpVv098cc,4361
|
|
525
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/egyhellaswag.py,sha256=44eoNICcU6-zYrmpUl4YGuJqSqBJ9KTew6Wzj-qfH-k,4577
|
|
526
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/egymmlu.py,sha256=Jtia9IzuQm_FU_3vk0o0BG7za_TNP_YvSiFk10J3IqA,6369
|
|
527
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/epec.py,sha256=8Cd2eZsUelZLNyJs8KDuo_d7oSHMtCLM7PV2OVp-mco,5713
|
|
528
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq.py,sha256=oZmCA7Pn6bLDWtLXBT3KYpgDgxJOnxa0hZctOmLt2hQ,4260
|
|
529
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench.py,sha256=1IkAkQc_7xF5CEnhXQhcIxfC_XZahIK4n3RnWK5vZ_I,8045
|
|
530
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_ca.py,sha256=4RgL2wIBPJEBb4ANJpD39qiFSeOkCeIGVXdrtMLIJQY,5988
|
|
531
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_es.py,sha256=sPqsKjSUkszF5sShH5AekB822TX9UCCfZbMR215fqR8,5988
|
|
532
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/esbbq.py,sha256=3lvIcWmqVO7K6CZHa9cEv66XzhGTLwoIBHQNGl1_VxA,5938
|
|
533
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/escola.py,sha256=cDZkmuv0TXe4ShayNiLxMnaXP3B_9W7cxAOb-z3UO2I,3121
|
|
534
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ethics.py,sha256=hlwNlzBsNU__nYknMvBtmEv4Z71vbmQC1Soq19Gfoac,5571
|
|
535
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ethos.py,sha256=TcDYU5eG0Ti4UumMVDt_RqjohJf9nESuxsXIhqyPCXU,3860
|
|
536
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus.py,sha256=gjzA8xMvYRWjuGWoVltXlmJgi4KuXlQclbIY4w7shPw,4264
|
|
537
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_exams.py,sha256=p4NGqMd8S5vi97JbFGihhvp6D0PV7gihn0amVfG0cA0,8731
|
|
538
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_proficiency.py,sha256=iiLuH0hUypAMPcIqw91FsBOby4zqB9j8fJoj6EHLgj8,6340
|
|
539
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_reading.py,sha256=IbR_RZLs4qM5YRf5RUOOHj4N-UEYYuy6cn653iyYFx4,6292
|
|
540
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_trivia.py,sha256=ppuPnMO9_niyG_z3I4K-wojCI3vif1jHGQLbmDQiJCk,6312
|
|
541
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/evalita_llm.py,sha256=DW_u-xNGkF_84EtmpveX49Umt1RVeqEdy7qMq3i92CE,6596
|
|
542
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/evalita_sp.py,sha256=rvmw3MJpPHln4p2FdA8b1n9lcd-x1wvBJcUALH5nHm8,3970
|
|
543
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/fda.py,sha256=WsQZGYDZUwb_Krhy9RjO6c1eSjIMdLif4osWB8eAq30,3944
|
|
544
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/financial.py,sha256=fkG64F2F8KBQDeG6BKd0ksWbl6RQnQkP9yeFNhzCRdI,4288
|
|
545
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/flan.py,sha256=MHFSgBXRp2vk9VW7IVYSy6wY64Ne1gzTa8kLyEddK0k,4577
|
|
546
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/fld.py,sha256=GA7Jlu4XtE0nbS0lGCjx50lt8eJk5ZlT3bTtFyUX-v4,5216
|
|
547
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench.py,sha256=L-Di_tpAd4KjcKnAeuEZ2JNjjVBQI9fqhMNaUFkeQZM,7820
|
|
548
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench_mc.py,sha256=H7BUHk857ioHCu9KhCjX6OTasagalE3_tk_Nw3Y1e2k,3616
|
|
549
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench_perplexity.py,sha256=ZWiytkj1TlCmKfouvpfJHz2dVmGMZ6ScDrOFXVcdPOA,3119
|
|
550
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galcola.py,sha256=0ORvzlWMGi-9ybDKWSPG5pDe74gPTyigiHlMcB1fG6w,3960
|
|
551
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench.py,sha256=KKGFFEmt0KGcFBtbZ1Usou62n6lAEuqIgoMrRn02HMI,6122
|
|
552
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench_gen.py,sha256=3-GQBmOcYU_AH4kkVuudC-0i3z8zUR7gGMFH6SU8M0E,3956
|
|
553
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench_mc.py,sha256=uuFVeFyNfEp-YUpNJshadahCvG8c5AAlr0ebxHAO7gc,4159
|
|
554
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gaokao.py,sha256=KqAiVV-i_0m9cbEfMQFHaB4GKMTxsjeWK_cVLwyq3hM,4847
|
|
555
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/glianorex.py,sha256=EggmYhwm8vfY_6CZgpPB0SKEAxaxUgBOSJgYzP0F9Kw,4565
|
|
556
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_mmlu.py,sha256=GB-PacZ0kk96e7jxGQK0hHuu9u9cWI6CQWbeM3DpdZ4,6444
|
|
557
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_piqa.py,sha256=aWrPCIW_pOPlfLXtRbzQZZkEcVO9A7z4bnLwztyLJyg,5990
|
|
558
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/glue.py,sha256=zIk_KN6wOCs-1ZBNvWr2VAGaRK2-0rX2pmFBQYWyIas,3739
|
|
559
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpqa.py,sha256=OnRcBqQks1j0d2wz-Lwv1kt1HRdzo9v8djF50yWxO2Q,5641
|
|
560
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpt3.py,sha256=jLlJFYXCiO20kWI22OO64rxWjp1W9kjjD13ygx9DNE0,4354
|
|
561
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/groundcocoa.py,sha256=x8PbYJ_TvyeEKj33JD-CfyuR23jzCVNUuEX-zvj-8t8,7246
|
|
562
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gsm.py,sha256=JlretZyrsYcmJ0KyPYFWGLYFx8Cvt3-TvhgUOjahEmc,3830
|
|
563
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gsm8k.py,sha256=p5CAbdyn6RRZ5I4jwIlB0keS0XQfcK_HTfvdQmLfI08,4603
|
|
564
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/haerae.py,sha256=4AUa0l3HgFqj-Lw5UvmRr6gV2b29ZgLyngLK7MS3nTE,5947
|
|
565
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/headqa.py,sha256=6XAzuoqSDahyjX5BjE23zTUlv3KN5_PHlu0fjRIsrs8,3875
|
|
566
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hellaswag.py,sha256=EFwXe4TAQsJtSq5U3I7yQK6_9GG_K2quG5PrQc-xJNo,4412
|
|
567
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_ethics.py,sha256=WA9FnU567x3h5tBfOFRJidDy9vTESCwq1-c-fhmSjvs,9149
|
|
568
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_math.py,sha256=TNDEF2Hu4VaXeQxpn0Pwt_0xKnpZNmjhhP8qxgehrKU,6657
|
|
569
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/histoires_morales.py,sha256=VbUKmBd9XAmvg2iJB9KG0UxXx_gEc9sFcwXkKDvuAcA,7236
|
|
570
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hle.py,sha256=l0PNHf3taiF-UhE5VO7_GFXwtD9ReR9h1B9-eJ-ukyA,3791
|
|
571
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hrm8k.py,sha256=DRbfEHav5FJwBgKL7bCv4uHe66MwwzwUtSitJe86-e4,7711
|
|
572
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/humaneval.py,sha256=FBNiL5EvqxbtnNurZDKAgiTb_P28_S0DVQN3QEKBGgY,4377
|
|
573
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/humaneval_infilling.py,sha256=DF0xtvaVnFD8BE1Y-NBgiVX22FxFRPHnxwL_ozxnNl0,6062
|
|
574
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/icelandic_winogrande.py,sha256=LAo6IvFinvL7-twDd5l1FK5Pa1Q99ZhgOETOV1dW9bI,6071
|
|
575
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ifeval.py,sha256=J7AHYgnEKuePzvIR9pmMsCoaJNs6QIaSIt5MHLIHjUo,4225
|
|
576
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse.py,sha256=Nq00uxvdezvBZhJKlL-T1GpAGxe-_r7PulU7kZ591vM,4280
|
|
577
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse_scaling.py,sha256=gd29ZSQ33KZ_LLQP2QW7lqRK_Yq-fesT4b2wyReFewI,7680
|
|
578
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/iwslt2017.py,sha256=noRk2FxmCjnk2RgGHWuvaoT3PqnjF9H-V1b8fogqKQ0,4444
|
|
579
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ja.py,sha256=3nOdcPHjGY5nvubXdGfuO5R9D6J261CQAqjvL2zBVgU,4260
|
|
580
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard.py,sha256=XFLJEMJ36osS60OUzciVMXC4rtu1CO47dJrB18YZwc0,6150
|
|
581
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard_gen.py,sha256=K3abgi4Mywn9BQMSQz_V4Hg8etbLHrBE-NpJa3LIivg,8857
|
|
582
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard_mc.py,sha256=eIVjIolhbqineaukBxU2sf2EusFy6EjJ62gwKuu1zBQ,5037
|
|
583
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/jsonschema_bench.py,sha256=q-smnHB3IBpuSUebVPkcZGCqwZc6UEGD7p4MpvVNW4A,4687
|
|
584
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kbl.py,sha256=ZTQ_2CF5Rj1xgU7bnZAVrDvv9F2sVORrCCmR9ZjOjEA,5179
|
|
585
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu.py,sha256=yqeFyZ0-o8nCbrv8b8zcat3QadREiYAPV5fbrJNpcQw,6651
|
|
586
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu_cot.py,sha256=CvAVl1xMmvKOhpMd5INKXGjZs7wolf-hyLD-oBsH6a4,3291
|
|
587
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu_mc.py,sha256=v2KK3DCmTkZyVmkw9gSTHElHV28AAp9Dji2vGizd_kE,4086
|
|
588
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kobest.py,sha256=G8yiAzdC_o-1JC81nTYKYTbD9rZoz1l-GHmW8Hdk8oc,6625
|
|
589
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kormedmcqa.py,sha256=LVjOwFoNedZVFqIPv0dubaonjUhdU5Hwh2uwYjG7CTM,5384
|
|
590
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada.py,sha256=kFbqrlO0SSn8QBhVvFd4SvoQzr79efmHBrVOBgDM9VA,4892
|
|
591
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_cloze.py,sha256=9-IVLeyeiVQ8WhZnRL2pFJLfQoQGgdRYSxsYSLy2Hik,7210
|
|
592
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_multilingual.py,sha256=RnEBqMWvbCHP3kCfNlJ__ymWu6ogwdJyuPHaIBjZ698,7297
|
|
593
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_multilingual_stablelm.py,sha256=S8IBwsTM5yRzE2RRfjF4XakO9QxcIF_vmLbUceh1v4A,5157
|
|
594
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/law.py,sha256=96Hrld3W3srvQa8b1qWYEBbZs8p1EU2-Mn0Q1kbguVE,4264
|
|
595
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/leaderboard.py,sha256=nPRQZ9996-Jc1yrnElMyebbyZdyUA9ssnG2t7nhUqbM,8075
|
|
596
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/libra.py,sha256=le4wRHUc2Dpq0gTzS_lQKCZ8zBvlYMVFq0OMkzkdrck,6107
|
|
597
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lingoly.py,sha256=nl_lf8-_pcNDSVbqkGaF1MPeiNYdnLa4H4qCm7rmBFk,7752
|
|
598
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/livemathbench.py,sha256=ZP66qtfYsyfZN9nyCSafTANSenetKMFA0jQowztxEbY,6217
|
|
599
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/llama3.py,sha256=RiQLmKvG24XtH7ryrquAexP2Bqc1z16WV3gmsurrcAM,5947
|
|
600
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lm_syneval.py,sha256=3N2Bt-aAaxugVmQGwt0MdQZ1qKzm8i_qWuqeQJ28zZ8,5981
|
|
601
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logieval.py,sha256=b8mWP5I0RUsmLSDE0UyYg9VC87C_qGQWn8Uq78G_dAE,2932
|
|
602
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa.py,sha256=tOAbI7oXs8MnsAT5VdPQuLzGfpwqzRfke_z0CDwLF20,3989
|
|
603
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa2.py,sha256=gRnnCQnDE9U3paCbsU0H1sjMC7Ubuil0E4bpBZLynC4,3900
|
|
604
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbench.py,sha256=5X73swbNvj7flqtFlNAqA-EGKX22QZM2mVeqtn2KW8Y,5974
|
|
605
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbenchv2.py,sha256=5f89b7cg4NHa_1vsIbFFz-osh3srP_4KMfbCzWsHc9E,5992
|
|
606
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mastermind.py,sha256=iyXStISnO5I9KBkgy9sh0nC9eAxmubZ4Iu0J8hIbga8,8019
|
|
607
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mathqa.py,sha256=Pd-E5D6pRWjN7J8rMVy3IbnU4DuEd-20tAyHrMkxiYc,4779
|
|
608
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mbpp.py,sha256=98X3RksIJ4GePIGiGMGN_44mxKm26P3n3hvkO6g0Vb4,4210
|
|
609
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mc-taco.py,sha256=rSvVod1RH3CWYzZi1hH4MJOnYvnFYFwhOSbwQhckSJk,3955
|
|
610
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/med_concepts_qa.py,sha256=QbCTKrOXGmd6oP7Pt-yRa8idy5ZcIROuXLsjsSeouRM,8785
|
|
611
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meddialog.py,sha256=ld97Bu1mJneVcQ1RlcEyJvnRwxLCJUfYReHKSlMwv9Q,7082
|
|
612
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medical.py,sha256=KzHsg8KVHTawJS7C5nLIYGw-LsFgq2ihr6NZ-bgFYDo,4280
|
|
613
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mediqa_qa2019.py,sha256=W34U00Bk3jwolmds8-P2X-3fSULK4Eo2k0hACs5Qf_8,4796
|
|
614
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medmcqa.py,sha256=_o3_lWxSjmF0vh_JqnjJDKRp2nkN_ukUGDziKwFebdo,6711
|
|
615
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medqa.py,sha256=J1ixlsjUUccQNCM-ShGaSTntBkBUX1lOObkV5a6Hnwo,4026
|
|
616
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medtext.py,sha256=OeKPSgBPfz7XjboeW3KGVmSq19xjAFp6dYz0tASwujA,4143
|
|
617
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mela.py,sha256=FXPKqXefxdMTftO3ob6Xx6Yetm9ZQOgENpHFsEOg5B4,3828
|
|
618
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meqsum.py,sha256=C1DZEn4peQm9GVfMs4fE0yKkMzwCXReZ88rsFzDSigk,4259
|
|
619
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/metabench.py,sha256=ooghyQ-N5KCtlCFtGP91V_hKQEkSD_fWjESu_GSxwFA,6041
|
|
620
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mgsm.py,sha256=K11m-ohBZ5Pxflp9ZaL8rXIeHbKmpHj2rX-uba72lcA,5060
|
|
621
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mimic_repsum.py,sha256=VjHLEhKwuHkHo8s8gnLg_5qR5XE3wdnizdzxQVGOG64,5351
|
|
622
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/minerva_math.py,sha256=UovLnmz4vU1mU9-HGBiH_WzPnXBPkJAhrg55mo2T3Xs,7213
|
|
623
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mlqa.py,sha256=VSx6TrBBLl125o4tF_m6o8e4huRInd8-dKs4lMkTdEA,5659
|
|
624
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu.py,sha256=J-iEuRvPmY3_ORP-d3l95TV2TWvo4Zqn81qW0zzzhOQ,5475
|
|
625
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_cot.py,sha256=5wtrR3gLHlhf6qy_yA7rQZOl6zxumrMbd59LqMSw7EM,3284
|
|
626
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_mc.py,sha256=MI_L-X1QK9OglTbxNYRhkL-nBRumVXGPk_nLRZpt700,4079
|
|
627
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_pro.py,sha256=A1xG34WxJmIkrRsGtFX_Bfr8rKFIjFl3kLlJnYplO9w,5353
|
|
628
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlusr.py,sha256=gJehtqOZkZtcPytpshr6j0d_rmBiHiZpOzWawySw_Fs,7988
|
|
629
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmmu.py,sha256=3Hh2-BdiHZpz528vPniN9f0Ac5dhuoo0J6_wRg2grN8,5606
|
|
630
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mnli.py,sha256=tZvZzsyp2Gxk18KEECx8uX2JLvB4VFKyJgTI0S6CGLQ,4241
|
|
631
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/model_written_evals.py,sha256=y-NGZVAl2AA0mN2kaFwZoNV6TouR6L-3kAWatld5q74,4591
|
|
632
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/moral_stories.py,sha256=cUK-aNA25X-hJsv415dz_J8I35ChcinYMYAPMbP4y08,5957
|
|
633
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mrpc.py,sha256=XMaAyhwhM7UCnf_rr474x7C42DXHpKftWdDlpYH7p_0,3820
|
|
634
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mts_dialog.py,sha256=x54oU2Wr1cg_7X1NYO-gBSV9iEbI9pRykEwR9Zlsg1o,4308
|
|
635
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mts_dialog_perplexity.py,sha256=o3ZsLcemHjPb2qCJlv7i7kFKLFkKKbRIIshjwmPTdcI,3622
|
|
636
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multiblimp.py,sha256=EbtO0PwvTIycRpTKNpayw5OB1wsj3MrghxIpHwdpd7s,6260
|
|
637
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multilingual.py,sha256=9VBd8kU21XX5P733a1qf8JkID1sdo6m7myZJJcq1kJc,3722
|
|
638
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multirc.py,sha256=QNkhgWHvfWnKTgQ9WqIYgpo8TwMo4C78MlaqRlzuRts,3929
|
|
639
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mutual.py,sha256=OaGST2AovGTnuiKj7dLL3Uoo1I0u-7UZoW3TmkVJwOI,3860
|
|
640
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/non.py,sha256=3Ruriwz_Tb4y1OroMaO8KWZmV9tEcB8MgQdU2N5uMU8,4264
|
|
641
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval.py,sha256=kZE4fRbg097KGSzyCF_bsOCDg4GooFAOTItWWpNUn0w,6973
|
|
642
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_exact.py,sha256=lKJF-H1lV8-_X4imnfWGs2IAYd_77yAbGvfmGAb5Egk,6140
|
|
643
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_gen.py,sha256=KROJgvdvAtd_Otrnss_EzPxF6upZiBmRRAYzvxJzVsE,11344
|
|
644
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_gen_exact.py,sha256=zi3V5pNcbIev9mPwOwhyXSATm9oOcvn3xMVG_MPW1P4,6472
|
|
645
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc.py,sha256=VDxUFDOhsnCvabkl4D6Fda4qfTDJyYun0jYcUj94DvM,9603
|
|
646
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc_log_likelihoods.py,sha256=mMVdFEJJ5vM63kC429DZzPbT5Bupp322ZuUYjpjHBco,9392
|
|
647
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noticia.py,sha256=fBvmsoMiGB-N5FZ3t39ZnTW44cKPIrXYpKZ-iBrWLjs,4896
|
|
648
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/nq_open.py,sha256=KZAuaRKcm6-askmTkmJzbl97CbzeB3rxOieIxGLVjBs,4632
|
|
649
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi.py,sha256=WiABcd4en7SjqkTsKq9ACWB_-tPULj6myzrjqNCevKk,864
|
|
650
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_arc_multilingual.py,sha256=JKBcUO1ItfE0IHxEbGvLlhZaEpwHJ6bqNmbUKzsd8rE,6835
|
|
651
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_hellaswag_multilingual.py,sha256=vvqHq24sVxjmQablTlR4zI_9JVp6msl8a8gSWwzha-Y,7424
|
|
652
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_mmlu_multilingual.py,sha256=UI6gwnQL8LwZmIxzatwVjF39CPs-UMdu3uzq1WFQmms,6618
|
|
653
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_truthfulqa_multilingual.py,sha256=m254nxIhoqRYlVXK4R5w2ynFqDdidPRKEYgvlfq_W9E,9505
|
|
654
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/olaph.py,sha256=NySRIqiWaiqJcGxoIDGo5Lefxkliz2fcrMSrPd8Hhfs,7279
|
|
655
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/olaph_perplexity.py,sha256=3aff3dr8ZHWw4_FRVqweVQNL16VmHokKlmsoTyBSuVc,3594
|
|
656
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/openbookqa.py,sha256=HDfNLMKVT-2mKOXzEUO7kBeMe_VxSHzhfM6VoiqyKJw,4108
|
|
657
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/option.py,sha256=xT2bpvNzLwHpv_VmiE25Pvh4VSCYvh5LGUMGFRnIchs,4276
|
|
658
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paloma.py,sha256=Y7vKEU051LHAtmzb7lAO7LgNm4orDmWLb0D1AVBxrHA,7904
|
|
659
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafraseja.py,sha256=fk65rmGff9JHwUVuF9PwdRlS40z_bTFS4YuUKsqOMEc,4366
|
|
660
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafrases.py,sha256=vu_ttWB_ZQolk81am8IuCiweK4qywNp3mZQJqvspSFE,4364
|
|
661
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws.py,sha256=RxPeMbRdaKfcIjKqeA-NyOyGoaZTL6k142cdrr9LDBI,4268
|
|
662
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws_x.py,sha256=LNpE0cZ_Ox4EBLgLWhjjoLdQ2UnmTNRL-q-8DRe_NnY,6009
|
|
663
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pawsx.py,sha256=YE34VO0dYrXLgSMbTpgif-HSM7uXoWJDQlud7H0nPc4,4055
|
|
664
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/persona.py,sha256=t_QLyi4v0wbXk13B2Fd1pJCDlVrAXT5RCOtuaZsIIWM,11055
|
|
665
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/phrases.py,sha256=E7I5eEpN6-Qc4bUDj1GzC8hO3cqatFrZ1oRDl0gF7GQ,6130
|
|
666
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/phrases_ca_va.py,sha256=F3p14GK8kpSsknnl88JYFY0ezuZmgWit8iel7QBtIc0,2986
|
|
667
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile.py,sha256=cfr9GrsLPX4XXAORV3PkXpC4_2hKb7u24PswVLAMnYc,6432
|
|
668
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile_10k.py,sha256=K5diNndk9obpncxS2h2pJ3WeFpceZTlCO7d1QXeKCYA,4845
|
|
669
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/piqa.py,sha256=iXN0CxFzLDvLt8pWyuV9MlOUXIiIIkN7obg79eNPtnE,3884
|
|
670
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/polemo2.py,sha256=8WYXV84sPbNLxwfnPLSVTDKY_suFlJPLp9ffbs9wnZ8,5179
|
|
671
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/polymath.py,sha256=r5KMsAeXAFsKKgINNmY6xUDXd8KchbxQA4kzo6ZUGDE,6190
|
|
672
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench.py,sha256=Jg3foz6stbTDqr9Fi2B1YUd2ktH_PRvfi3-VfCazs98,6105
|
|
673
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench_gen.py,sha256=hGLtogrXR_yib9riVQZAkqpwCsMbhb7F8K-AR0uKFgU,4191
|
|
674
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench_mc.py,sha256=u02jIQXUOuwbjfCrDtd1rIEvPq4e9CPvvsa7De8DSYk,4133
|
|
675
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prompt.py,sha256=DBUw7iKa56bniQjGOYi81q0fusND2A6fVufeFpFbjkE,4276
|
|
676
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prost.py,sha256=-Ub_F46PhmsBCCP2r4t_o4LpPnlyBAWljcBfGzoCn-M,4060
|
|
677
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pubmedqa.py,sha256=HKZ07Ysvyhj3nqGS0LgtYD0IB4vRzjHMC5420FRpzB4,3971
|
|
678
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qa4mre.py,sha256=5HUAPE93IX-0wbkNBpg7WZI27u2Ki2snF8BgbpZ8cDs,4154
|
|
679
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper.py,sha256=IqngDM_SscCLYfoh9IlPGuJqGdL7MByIvFMBUF7ZPHQ,3961
|
|
680
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper_bool.py,sha256=b_aQQMVgr8GcMuTmRemXYbvcJ7bTCCHe-lF9g-s8jPo,3839
|
|
681
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnli.py,sha256=t1OUysTdpLEawtGPNdPLu88P2PRcum0BuC6webyC1M0,3797
|
|
682
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnlieu.py,sha256=Ku4MpOlfBOsnP5utbA78UecsYpoL67C_jqv0EOEtr_Q,4276
|
|
683
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qqp.py,sha256=qw1E7XxC5rOfDNWDysrvKxZkJ6k4iaw0OVEg1AmMCnk,3796
|
|
684
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/quac.py,sha256=M7HVhIic8RB3LvV0LCsDdMqEZCPBK7sWb_fkH-Y8HGY,3797
|
|
685
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/race.py,sha256=zjZ5liczuQiSSwYFJWtpcgaxEj3Fo8kEf6gR6hMcSOQ,4141
|
|
686
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/random.py,sha256=krT116DEDmYHd9D6ym2RxswrUD9FWOtaYuEVS75Wj0w,4276
|
|
687
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/realtoxicityprompts.py,sha256=1r13H8HW4N4J9eSDmFCAmdoKGhShjxjZ1PCJzhSLTcg,4575
|
|
688
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/record.py,sha256=Rq2W6-7Wq3HhvPkY_psn5rEQJ5wXWwvxzZx8gevQupM,4261
|
|
689
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/reversed.py,sha256=wyqkUVJKm1Z4a08iRtyWqc8oZu-b7-uRCuPlnNqGSBk,4353
|
|
690
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/rte.py,sha256=zv5RTqERoz0VrxHMD4BldP5K7o9CZNq6gqqqQRjarLc,3785
|
|
691
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ruler.py,sha256=7dCa8wZ1QaC2OjQ8XSyZYj4KxVnbQgI48oL4a6VdZHk,6829
|
|
692
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sciq.py,sha256=qq9Bns5Luby1FgoHQ8E_zpWXXiLji8iVtcO112426Ic,3907
|
|
693
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/score.py,sha256=gVsOLD3_elWMEJpcnuBC834HSW7Qrdgo8tzCpIcLfDg,6787
|
|
694
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls.py,sha256=EBWbrmRFYp0dh2Uu2nEFzivc6g7jV1nNPJ0jhaGhpBg,6131
|
|
695
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls_mc.py,sha256=k0TBC17INJMgPlzw3oP6B89kN8Z6WQjvz-GI_FU_NQ0,6042
|
|
696
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/self.py,sha256=wo40Mk6a3mmR3vnVEhMMyMh_Z-lvQtJHhsTFpxc2vtk,4343
|
|
697
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue.py,sha256=Cv4UwWLmmHjgghMRKjVDCX_NLahZlwEYQEP5ctv9fzI,5368
|
|
698
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue_rte.py,sha256=r2YKxLs5c1d_SAwuvKkpgY8Xme-_BrMSPRi51g6w6uM,4180
|
|
699
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/simple_cooccurrence_bias.py,sha256=pipWwa2PNQVTfJeqU1xi_lPbv1XVUdgedKXsku2Lrpo,4490
|
|
700
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/siqa.py,sha256=9sB7reIkFAvekmnTdqS5jML3dTMrePkTY4YyaN97Qvk,8229
|
|
701
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/social_iqa.py,sha256=r5BT7cy-1-N-qhwBJGERDXtok9UqTGyZ_v69FiTOJTA,4055
|
|
702
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench.py,sha256=KD_S6j-6s0VTU4JoZpJWAePQe8jX-sf61GwHbGY8q_A,6075
|
|
703
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench_gen.py,sha256=N878fSFoeO3334gnht8zg5Gfb_sy1lbsZcsj0D5ENck,3919
|
|
704
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench_mc.py,sha256=izYHAzjLmdFVfmUMB21ZZZhiUtDIMUdyKo9yNfPm404,4128
|
|
705
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/squad2.py,sha256=vMQezYU6gN0aD5oiLItMCjRAxxXa1_GBYc6VBCSeJMs,4602
|
|
706
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/squad_completion.py,sha256=qrkVhWAw-gKGfwNCqglqFP1TCU1_LPJQmIr6qd2i2iQ,4410
|
|
707
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sst2.py,sha256=eqpA1VNI08nMdxF4ltNq333RJLICvENxSS7svQY57Z4,3851
|
|
708
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/storycloze.py,sha256=NVsKcygSoSvFPK8sR7OESS7Xo77fTtDP7zWp-tCRA9I,9610
|
|
709
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/summarization.py,sha256=fN0DBEbhmoddKzTLiQfeZepxC3BXQJ8g4BpVZoHEAYQ,4304
|
|
710
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super.py,sha256=sU1s68McNYfsNOOTy5awKECOfUSQTVloaBPDlrq1HVQ,4272
|
|
711
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super_glue.py,sha256=7VKmF7tz6b8NKsqyHUM0HtciO47AHntQlBfI2tLfRfE,6049
|
|
712
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/superglue.py,sha256=NosWWhOPRF45WkNCgJFJXI4x132tFwRX102DcdrXf0E,3827
|
|
713
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/supergpqa.py,sha256=4hIJpmgdVxZf2xsFOkbvgIkfyFAXF0vbnwlCNyoYNJU,3827
|
|
714
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swag.py,sha256=2Dx3mNKV62o2hNSqzdLXc3Oa0A6ddxEBCAZxMAy6WYY,3946
|
|
715
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swde.py,sha256=uE8falrDdx6_mAIAHRAYmqieQPATtiyVBu0n8Gu3pTM,6978
|
|
716
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sycophancy.py,sha256=IqJcLdsXNIbjIZKbELv5r6_XWKFUXTtNnbzpwHzla1Q,4375
|
|
717
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/t0.py,sha256=F3V7vUCS4-U1_OS_yQPxtz2E3yGTckJHUKjW4cyHKCc,4321
|
|
718
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/teca.py,sha256=DSeFGcbi3AzZ6svgHeUZmkaGOIgh1GwfOuWRsOR7R5U,4331
|
|
719
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyarc.py,sha256=dxHaJWhITTunfEOZ4ko9rWMzSo905lc4aMpIMBWXr14,4346
|
|
720
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinybenchmarks.py,sha256=N1OIJjW4FCC4nk97PUB1Snf_SQrkSOBI5S7KGIqnOXk,6092
|
|
721
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinygsm8k.py,sha256=EoVAdXegybLiH3alutR5Mwu6v_M7QtshRGUtsBrhJ4k,4351
|
|
722
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyhellaswag.py,sha256=IBscRNFLP6Lw4GdzEDUnF99hdjRB6gs0PqQiC0yCluk,4376
|
|
723
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinymmlu.py,sha256=8kHFHNkXAEaOC_vCsaikK2VPPWBIMzV7Q3Abq8aOZgQ,4351
|
|
724
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinytruthfulqa.py,sha256=UyYnf5BrrVG6RmqqQO00FpZbs048bzqi3tXVRjHY1bU,4413
|
|
725
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinywinogrande.py,sha256=aHPaAnzUNJf4hP2CHR1tgx4cmQpDMIW5TyZGVY3HIPE,4381
|
|
726
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tmmluplus.py,sha256=Om7XEfXXiVUt6fBt9pxgzB-rq58a8HwEQIrt1uR3iS8,8206
|
|
727
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/toxigen.py,sha256=wyr9CDq4LupITsRR327vGhIfcgjXdtDgcZBPS0EexDc,3270
|
|
728
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/translation.py,sha256=rHNpXKyLCl5qUtTw0nMjg0jW4sFOK9SOKkawc0UBH3g,5167
|
|
729
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/triviaqa.py,sha256=-R3_JDrAIDk68ejFptAeESvsx_OxTNbTNEmke0mFnLw,4414
|
|
730
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa.py,sha256=QE3LhA1lDiP-DnYBceul4p5mDc5uUmml7dXXu4aKXdQ,4480
|
|
731
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc1.py,sha256=orxcmYBJkZD0SmH_9tHULNXTxe8VVFq0tQE2cF-a3fA,4102
|
|
732
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc2.py,sha256=bgMmg2nHCk_t-Z_KsgcuQ8jc9JiRhiakNyzbm3z0bYw,5096
|
|
733
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_multi.py,sha256=oCWC6hY3RI9Nj04Rij71fDZCT-L17kXirkGE9fyob-I,4945
|
|
734
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turblimp_core.py,sha256=L1T6pDwdy8s14F3O-cSxd5XE6rGU3qVpEkV1JiuWjJU,6008
|
|
735
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu.py,sha256=C6_VNlabW0JUH55Ls8Q3gu9QvrlRsxvZlLtZCSxxDC8,6338
|
|
736
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu_cot.py,sha256=nD7B0tA1rMd80eazkrnGjXS1VemkEp-pvqMweyGMphg,3871
|
|
737
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu_mc.py,sha256=DSkDuI1WXm7KAUUOkE86O9_NMcQHwIpnHQN6iUQSC9Q,3790
|
|
738
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/twenty_newsgroups.py,sha256=_mAy170d7m1SlPtbNNkToyr5zFasvyEUlWgJyuNZK38,4563
|
|
739
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/unitxt.py,sha256=OhhhQJfwBTSjWipi3nPIelpjrSpvI3vCZ1FmYPH5WGc,4592
|
|
740
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/unscramble.py,sha256=B7B9-pXNQsLlwqrh_CSBSHyqdGfhUgTzrEkj6FzIPf8,6048
|
|
741
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/vaxx.py,sha256=kJ_ZoftHm-qkwXlk3nIexFzdJtuY1UywNp4haNg9RSg,3581
|
|
742
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/webqs.py,sha256=qX3-nvQzZ4VV72bS2AUuBvgHnZEpwZGxz71_gUSAqOY,4438
|
|
743
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wic.py,sha256=vGI2GxJb6vfGJssxMNKMngprAI0P4OMDdoE2DJdpWG8,4108
|
|
744
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wikitext.py,sha256=r2izxF9w-LsQg3erIk3v8vGQttc7nxpqi0KHkYe_cPc,5000
|
|
745
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/winogender.py,sha256=1CoZe0Btrt06gu6TCfcPnobbewPgY2EOdijKISq0bgE,4766
|
|
746
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/winogrande.py,sha256=rynWcZsS17eb3mpGU7PifpRPzsJbhFeOKL5-0g3dC00,4114
|
|
747
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmdp.py,sha256=1CTyt1f2ABCsXLtKgobwWB0_89q_6Sb7b_06Km7EqEA,6030
|
|
748
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmt14.py,sha256=_zHPAk_V9VyigtVcpTLpbLgFoNrHeASloddh9lQLUGU,4043
|
|
749
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmt16.py,sha256=pX8F91Um3kUbsKgxJZDdRBfFzlmCr5WUx-OFBYGF5CM,4446
|
|
750
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wnli.py,sha256=kR3M4BF6AyVOYmY3nXdwHQHYLwXkZJrulrB5FAJ5FQ4,3855
|
|
751
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc.py,sha256=FTqa0mnmFfpuSVUOuLVosvBJOMpBszPiiGUvUybKLxM,3974
|
|
752
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc273.py,sha256=Q-xGA95r9l3qLV-PKBdXaLrkD6AIQmI5KU21ZrKJ9EM,7074
|
|
753
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xcopa.py,sha256=9c6CfB4wgdkwroe-RzOfudR3Zhcp7wF4HnEpvI7yaZs,7463
|
|
754
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xlsum.py,sha256=qlh0yIpvpcVBgS9-VXTcmVGyZf4wpiAXv3s6ELdZFr0,5371
|
|
755
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xnli.py,sha256=kryrRr9VJYaioi-c5cMg-kbF4l29Q-fqW0Ri2HkHIC8,4181
|
|
756
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xquad.py,sha256=GE1qp2Pc_FDTYFCt-rSRWi6u_b6hXQvvH1n4l26eZ-g,7745
|
|
757
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xstorycloze.py,sha256=11GknEiqca4grspitZFZRYllt2IFLqL9mWW3ZIyp-Bc,4530
|
|
758
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xwinograd.py,sha256=FO-LW5iMLNO3SMk2EKtWVyTWrpNSOsqum7qp0MqNHgI,4108
|
|
759
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/yahoo.py,sha256=oI2VKhoGTIVPrJGv4L7-o2H-8CauckVDql1wRMs_WWw,3949
|
|
760
|
+
wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/zhoblimp.py,sha256=JGAocj_5n2zJJ2H3Z9aqB7P43hRUoykf0XJNykqD9aY,6032
|
|
761
|
+
wisent/core/data_loaders/__init__.py,sha256=VAcOq6wRL3oihqeJz6Pk23u1enwEkM8w2lShv0V8Sik,8667
|
|
762
|
+
wisent/core/data_loaders/rotator.py,sha256=hlahoJz_cvl7pxsFwVDcF1fc9Uj690rb7aWpqdfq9HU,3794
|
|
763
|
+
wisent/core/data_loaders/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
764
|
+
wisent/core/data_loaders/core/atoms.py,sha256=26vJ1B35_8EoQrFOEdNvXHcJRGxSfqsYpFIe6zQ3ZTQ,3591
|
|
765
|
+
wisent/core/data_loaders/loaders/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
766
|
+
wisent/core/data_loaders/loaders/custom.py,sha256=Xe1sOHH3_dRjBnQg9vwMM-XA8ROn65dUr9TeT-nuNtQ,4144
|
|
767
|
+
wisent/core/data_loaders/loaders/huggingface_loader.py,sha256=AvMG2DpEAOMfzz339WlDJ3qeqLjvcqY7P4P7lVnjKTE,5147
|
|
768
|
+
wisent/core/data_loaders/loaders/lm_loader.py,sha256=GZyawEOXVMaJSPuD8nkGhNsmcWJ1XS3WzfXBAEQk8d0,62200
|
|
769
|
+
wisent/core/data_loaders/loaders/lm_loader_special_cases.py,sha256=-Ku7dOZIBgJjbNAbP5ihXvRn5KCStzSmk8_nLsshmzs,18885
|
|
770
|
+
wisent/core/data_loaders/loaders/task_interface_loader.py,sha256=QGDfqMvdTFfogF9P533b-Gw52uHe0GnsS64AfnVVlek,11771
|
|
771
|
+
wisent/core/errors/__init__.py,sha256=DX-T4zW2PwNvTcyXyt1T_-oZ_xXl4WhzqaOZKiDaWFo,5535
|
|
772
|
+
wisent/core/errors/error_codes.py,sha256=5VqOHG-BFA4Ueg17Mrb5fSwQ2mohVzTuFEAsgT_e0MY,31961
|
|
773
|
+
wisent/core/errors/error_handler.py,sha256=6OE7pQHOrOm6ZtSYlrN3GjyeWgCzkTF1Znsl0nd2Yog,4435
|
|
774
|
+
wisent/core/evaluators/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
775
|
+
wisent/core/evaluators/rotator.py,sha256=uIpWvA4okAvMAgZYNPnmL79n3MKilntvEAtq0OoUIYw,7657
|
|
776
|
+
wisent/core/evaluators/steering_evaluators.py,sha256=3b2voLO9BzRzNmEjhiw2GzOuHLyaEDFtuNzX3ZR1G_Q,14937
|
|
777
|
+
wisent/core/evaluators/synthetic_evaluator.py,sha256=odKzYAHp1pIIvZtvRpodfVbnCK_E_LEMg7Z4TJG2O9I,14284
|
|
778
|
+
wisent/core/evaluators/benchmark_specific/__init__.py,sha256=hTcHcv3GEQQWZ7Z_puSUbI8LlQuzljr_-GtMA19wiro,1432
|
|
779
|
+
wisent/core/evaluators/benchmark_specific/aime_evaluator.py,sha256=3VD6C0MsEyIO7i744QmSoZk5uuqE_OnhvkhG8hweKzk,3062
|
|
780
|
+
wisent/core/evaluators/benchmark_specific/conala_evaluator.py,sha256=gs6FrmyxU3DTIjtukGu4CjREzW5asQohQc-HHw6ihf4,10681
|
|
781
|
+
wisent/core/evaluators/benchmark_specific/exact_match_evaluator.py,sha256=rc-QfiAYOzAkd5MAfghj7kQ6pEEYFtr4630pt-HRXmE,2756
|
|
782
|
+
wisent/core/evaluators/benchmark_specific/f1_evaluator.py,sha256=6_9_U3mUKBSBpOFq7HGXnN3xrmGOSV-H2agTE_s1I_g,6608
|
|
783
|
+
wisent/core/evaluators/benchmark_specific/generation_evaluator.py,sha256=6XNmp774qI9PqzybqQEwKgI_gn3ui3mIZc71GT5JXcY,20052
|
|
784
|
+
wisent/core/evaluators/benchmark_specific/livemathbench_evaluator.py,sha256=qJaRVtdMJr-Of0LrtLkX8aVeWAoEQHHqTisrp4SsItI,15339
|
|
785
|
+
wisent/core/evaluators/benchmark_specific/log_likelihoods_evaluator.py,sha256=zKY8BJvb887By9oWkq7x1tk8BhWzpcILRH63wypWweg,8139
|
|
786
|
+
wisent/core/evaluators/benchmark_specific/math_evaluator.py,sha256=EP9EkMshztj16mhIvnEaU-U2dRibsAYJA0VMAkcB-Io,4375
|
|
787
|
+
wisent/core/evaluators/benchmark_specific/perplexity_evaluator.py,sha256=Ir1kyFycr3Fq-ugiUlO3OkzSwhXB_rc-qUCzNomrqDg,6497
|
|
788
|
+
wisent/core/evaluators/benchmark_specific/polymath_evaluator.py,sha256=bZfGRp_adPYcSmC6MXjTRPpc8hkYuy2BXi1WUMZQ954,5061
|
|
789
|
+
wisent/core/evaluators/benchmark_specific/coding/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
790
|
+
wisent/core/evaluators/benchmark_specific/coding/solution_generator.py,sha256=PX3RdwhrrYG_2nx10GVvQjdq-_6AFP-Q-yEGFA9rYpc,9329
|
|
791
|
+
wisent/core/evaluators/benchmark_specific/coding/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
792
|
+
wisent/core/evaluators/benchmark_specific/coding/metrics/evaluator.py,sha256=f6JlcZ6yaqxbh2kzdIsEuBhbHKBm1MbpYUPauRLhBx8,15091
|
|
793
|
+
wisent/core/evaluators/benchmark_specific/coding/metrics/passk.py,sha256=WCkdt9a1KG2qoHNQtXdKPupVNQtppPyUGeq5yF6ZO7M,2059
|
|
794
|
+
wisent/core/evaluators/benchmark_specific/coding/metrics/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
795
|
+
wisent/core/evaluators/benchmark_specific/coding/metrics/core/atoms.py,sha256=I4wwOUTDjJ_4Jo6Bbpl8Tkv0Uz9zpwNjzeWBrXru-9g,1001
|
|
796
|
+
wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
797
|
+
wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/cpp_sanitizer.py,sha256=PVFARxEo0kkPznrRPM6j-SXRlX2xAuo8MfN5zXFW41s,3504
|
|
798
|
+
wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/java_sanitizer.py,sha256=JUgCo2ZVI0d8lPYHuxrZSJffnN1uPjQQA-Z6gvvErmo,4550
|
|
799
|
+
wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/python_sanitizer.py,sha256=ZyqnoY-lTiOKzdCX1LZdRhrNwv4LJyf75yidylvC5bE,4921
|
|
800
|
+
wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/utils.py,sha256=qhmqMbSyNgJWDXLd0Myuqtr2AssdBVt-jLia6wseYQc,4043
|
|
801
|
+
wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
802
|
+
wisent/core/evaluators/benchmark_specific/coding/output_sanitizer/core/atoms.py,sha256=yoPxZg-jA9XE065erGV99Dj6aE3dFvKzogxLo_IWvq8,1185
|
|
803
|
+
wisent/core/evaluators/benchmark_specific/coding/providers/__init__.py,sha256=rpGxV7u93zVQdkXfywxPeq_L5jXNFc0cN1Zj0TsXw2E,731
|
|
804
|
+
wisent/core/evaluators/benchmark_specific/coding/providers/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
805
|
+
wisent/core/evaluators/benchmark_specific/coding/providers/core/atoms.py,sha256=6THxuLLpg2GRiKT82ZItyQWgLGx7_Ljzk4YhgjV8zfM,1090
|
|
806
|
+
wisent/core/evaluators/benchmark_specific/coding/providers/livecodebench/__init__.py,sha256=JiYrBGxEJ9uc9HLTDPuLHmdIrgTNlThjl--Nj8XyMfA,81
|
|
807
|
+
wisent/core/evaluators/benchmark_specific/coding/providers/livecodebench/provider.py,sha256=VWwUTBA6PDk_a3qA_6J5Kof05VAnxDeJD7pgYrSrXqA,9677
|
|
808
|
+
wisent/core/evaluators/benchmark_specific/coding/safe_docker/Dockerfile,sha256=G_cF13BLPsMyw7MCpSN1YsM9vslkHCskTNlaAwoBs7o,1076
|
|
809
|
+
wisent/core/evaluators/benchmark_specific/coding/safe_docker/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
810
|
+
wisent/core/evaluators/benchmark_specific/coding/safe_docker/entrypoint.py,sha256=hwZ1cH6cnBIXjwzIJxslThCCagbwqz_tSo5g7XfYgNk,4202
|
|
811
|
+
wisent/core/evaluators/benchmark_specific/coding/safe_docker/recipes.py,sha256=fGC477jAvoYoOmbCvpMRwfpqDll8I5KSANjTcMbN05s,2251
|
|
812
|
+
wisent/core/evaluators/benchmark_specific/coding/safe_docker/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
813
|
+
wisent/core/evaluators/benchmark_specific/coding/safe_docker/core/atoms.py,sha256=JE2eOKtiXGLNGSNqaPwtpPbcipoPeDdmVNxgt5_EHYE,3026
|
|
814
|
+
wisent/core/evaluators/benchmark_specific/coding/safe_docker/core/runtime.py,sha256=XpkOVGb1nxVH-D_zp4W9oVJAjYjvstsbq0g_ZG4dr2w,5596
|
|
815
|
+
wisent/core/evaluators/benchmark_specific/math_parsing/__init__.py,sha256=pnL0726_u3LuBWNGWiXFkkSAlKvUKQHkp8yjBJ8mM6k,50
|
|
816
|
+
wisent/core/evaluators/benchmark_specific/math_parsing/core.py,sha256=QbBAV-W0nxs_D-n7B2aoX1cb9EphtGE_ml8r-cP9AcU,52242
|
|
817
|
+
wisent/core/evaluators/benchmark_specific/math_parsing/extract_boxed.py,sha256=ucDkS8afAV4IWB92uoT_y57i7lbkKan5KgiEHamf_6Q,1226
|
|
818
|
+
wisent/core/evaluators/benchmark_specific/math_parsing/is_equiv.py,sha256=bt9izpChlBoK9cP7RfvGXSciRzc8M66Ia9kuLgkutSs,4622
|
|
819
|
+
wisent/core/evaluators/benchmark_specific/math_parsing/scripts.py,sha256=c3F1qW4E8-jl-poVpUhtALquU2PSjdbIxFh5Z_oufL0,26213
|
|
820
|
+
wisent/core/evaluators/core/__init__.py,sha256=j-B9kTyEWXKUz50fPvTzwEzS47InlFgDrwv6eBmCQwM,156
|
|
821
|
+
wisent/core/evaluators/core/atoms.py,sha256=0-mTetvq6dnypuFJoi7OB829M3MloGHXAxqiyCTrIZ0,5271
|
|
822
|
+
wisent/core/evaluators/custom/__init__.py,sha256=SLvWqLZFCQBDYeSlyASAqX4-CorQ8wsroMA6nonECp4,480
|
|
823
|
+
wisent/core/evaluators/custom/custom_evaluator.py,sha256=502Gfrt4FMfwgRolPMCwbzkM4G1mD4DsaBv_zX_lW_U,13236
|
|
824
|
+
wisent/core/evaluators/custom/examples/__init__.py,sha256=Jho4mJgB6eGc5CaRjtcyrOIZAxY5Z4lLqbPtEXc4SIs,1043
|
|
825
|
+
wisent/core/evaluators/custom/examples/desklib_detector.py,sha256=1ysygPE-TpO7sKqOEZtOvdGMv5nLp6n57MawU7mH3Ic,5566
|
|
826
|
+
wisent/core/evaluators/custom/examples/gptzero.py,sha256=El4Sypuee0a6QMv7HrPhfqfTeJbaEXQRLjuV_clPlgg,6129
|
|
827
|
+
wisent/core/evaluators/custom/examples/humanization.py,sha256=bNTDLNqckkCgq_T7elG5EdqBsidWzRh-ncbCKNibvMo,2677
|
|
828
|
+
wisent/core/evaluators/custom/examples/humanization_coherent.py,sha256=FXuRbSZ1NhwjGkC8nKKs6Ax-YkrYzSGuJqvBcWh5VWw,4641
|
|
829
|
+
wisent/core/evaluators/custom/examples/roberta_detector.py,sha256=mq9MHdm6sosuzV8qRt-rj6gaYFaWsodgfBt1ZpCqwFM,6107
|
|
830
|
+
wisent/core/evaluators/oracles/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
831
|
+
wisent/core/evaluators/oracles/interactive.py,sha256=f3v2_N17fKzGyeOxONRJbrbn8i5uMeZmRvMmF0ShNf0,2638
|
|
832
|
+
wisent/core/evaluators/oracles/nlp_evaluator.py,sha256=KxbnF-I2IFbBQpoYyjQKGbYh4NErsEuhTCRYX_Tob8o,18220
|
|
833
|
+
wisent/core/evaluators/oracles/truthfulqa_gen_evaluator.py,sha256=aGtcvf1dkLit3P0Y5bjhBGeSy0rn-ImGgKV2X_nAlao,6613
|
|
834
|
+
wisent/core/evaluators/oracles/user_specified.py,sha256=V1dKrNj3Oq7UC_I7DT0WGnktP7R_DSW6UAwDdrA8SnE,2360
|
|
835
|
+
wisent/core/evaluators/personalization/__init__.py,sha256=t3lWkNFXA27eDzEQIZnkpasjGJ0w6UaZee4dJEQpFEQ,311
|
|
836
|
+
wisent/core/evaluators/personalization/alignment.py,sha256=UcF5ismNElPbWkJNb-EuUKZraYE5aI46SAXoniQxQwE,5449
|
|
837
|
+
wisent/core/evaluators/personalization/coherence.py,sha256=PM6_m9au2jntw6z1IjsNurl1oH8VfvPrwJZLJROXZiQ,11582
|
|
838
|
+
wisent/core/evaluators/personalization/difference.py,sha256=8D9azBdqciYo4IJt2lwrFGrwpm7q3Kyze6K_yBYPHoo,2649
|
|
839
|
+
wisent/core/modalities/__init__.py,sha256=jtowcmTMQ_kOYai7FKwXY0vqSYQ8vews84VSMH73XFM,20138
|
|
840
|
+
wisent/core/models/__init__.py,sha256=G9JfaVU_CXCG_wnz5D5gjDJ-_G7zqHFmjaixDpbCixM,456
|
|
841
|
+
wisent/core/models/inference_config.py,sha256=0n8647avNb4w0o0UsbWnU_eh6cMU23ppdJ-FDYx8mAY,3704
|
|
842
|
+
wisent/core/models/wisent_model.py,sha256=TvE1ZmSVCKwN5gs8c5HX06ABGsCw7G9RMzEer75ilvw,36518
|
|
843
|
+
wisent/core/models/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
844
|
+
wisent/core/models/core/atoms.py,sha256=7QMaJHl6axH6Iy0woZFsgEsBga1IaMhVq7q7JoASRic,16107
|
|
845
|
+
wisent/core/opti/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
846
|
+
wisent/core/opti/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
847
|
+
wisent/core/opti/core/atoms.py,sha256=7D6TxhEcBKxCplajICCfpZQ0g2rsKB6iWoRB0Mwr7yg,5438
|
|
848
|
+
wisent/core/opti/methods/__init__.py,sha256=AJmvQZqWRAJ_EYzrF28uA1mN37C-AtVnUgCKfXaD0qQ,365
|
|
849
|
+
wisent/core/opti/methods/opti_classificator.py,sha256=qVnpA57ncSsPLOby-rXvEOI_c_cRp06pJSsOLtO-tuw,6864
|
|
850
|
+
wisent/core/opti/methods/opti_steering.py,sha256=Lr8dYKpqgGvuJauvYl7FNUMtR1cUAl93aepiqPIeslw,5407
|
|
851
|
+
wisent/core/opti/methods/opti_weights.py,sha256=cfJziHJqd7P43sbZFDLqXmlyW-po-7yNmfKkCp6LBzc,19656
|
|
852
|
+
wisent/core/optuna/__init__.py,sha256=efbr9QeNaeWMle9pppWBJXs5ObiPti9QZBQGBvuli-s,2250
|
|
853
|
+
wisent/core/optuna/classifier/__init__.py,sha256=vv2wCAbw8ccZxq3vxrQt9plUdbr7VJj-t2rRnh6jBR8,819
|
|
854
|
+
wisent/core/optuna/classifier/activation_generator.py,sha256=QB4Me1DlzmjN9C4En9tpfgV9K_qeO9Iqo7igTJw8LyA,14448
|
|
855
|
+
wisent/core/optuna/classifier/classifier_cache.py,sha256=Q8PNBJ2rpHWvAQRVPR_ax0zvnrbRKN0yUjOhCwXf9jw,17439
|
|
856
|
+
wisent/core/optuna/classifier/optuna_classifier_optimizer.py,sha256=T1K6CSWreQeX4X-AeAT0jaf5hBDp1bP92LN_tDzm500,25490
|
|
857
|
+
wisent/core/optuna/steering/__init__.py,sha256=C1O9Y2AMFsFvLq6fmTW_wJ5vnRXKVGftykRw3NaiHwM,508
|
|
858
|
+
wisent/core/optuna/steering/bigcode_evaluator_wrapper.py,sha256=ZVp159jbcdUAf8A2q5AwINak0n6YqBUuq0iDLvXObyQ,7157
|
|
859
|
+
wisent/core/optuna/steering/data_utils.py,sha256=PUQ8Gqu8oPfoGGFi_lTfaPXhTZ5MRQO2Gf73vRa7NDU,12431
|
|
860
|
+
wisent/core/optuna/steering/metrics.py,sha256=6UbW9AMrRpT1pgvDVK9rb1-ieqqOq1I1KdVh6MtB4CI,16394
|
|
861
|
+
wisent/core/optuna/steering/steering_optimization.py,sha256=w-wt0AUNBr5RmYvg58pepx2KDqQAjlM9FBU2a2OgaRk,44893
|
|
862
|
+
wisent/core/parser_arguments/__init__.py,sha256=MpBsnsPcZQ8cqAn1tN4OYfb3oMQraNW-23g3Z5gSqeg,299
|
|
863
|
+
wisent/core/parser_arguments/agent_parser.py,sha256=PG3V-F9xKE53guiiQ2GudGKZnjvtoAbdyMI8Qhm_sZw,4971
|
|
864
|
+
wisent/core/parser_arguments/check_linearity_parser.py,sha256=HnYdSYOOP9MvVyRjO2ZvTIlKU-y_Hw3yXL5F_ehjZXs,2013
|
|
865
|
+
wisent/core/parser_arguments/configure_model_parser.py,sha256=LmsS0mYF7eP-zpYB7SdEbtFZLVtPtcbkEi9CMkYqMVo,350
|
|
866
|
+
wisent/core/parser_arguments/create_steering_vector_parser.py,sha256=GvyHt0fqwVk5dDmAsilOttsybZ3jj5Z6dQWzE8EUj-Y,1735
|
|
867
|
+
wisent/core/parser_arguments/diagnose_pairs_parser.py,sha256=i_kV0sEYkpWEur0BK32Y3nMMRwjz7bAjUY9BbhqSDws,591
|
|
868
|
+
wisent/core/parser_arguments/diagnose_vectors_parser.py,sha256=NGQGSPqUwiMLW8hwglWNPk4dSsL29374X2l-aiZjUZI,1824
|
|
869
|
+
wisent/core/parser_arguments/evaluate_parser.py,sha256=XcPdhNIHq4i6BHyvKxWeSJp8IuB3iihqzmhPFhH0VMM,1763
|
|
870
|
+
wisent/core/parser_arguments/evaluate_refusal_parser.py,sha256=UAqio0QQpoaNVp5sSg9On0H_rBuIArmOrMw26MQ2rT4,1534
|
|
871
|
+
wisent/core/parser_arguments/evaluate_responses_parser.py,sha256=Cjrjet4j0L8KvLEsFZxd4VcGNwuVn3PXoCa3gdpgNEM,905
|
|
872
|
+
wisent/core/parser_arguments/full_optimize_parser.py,sha256=BTd2Wc_qUkrdwvaPeoXkFogWB8ggr9q_kYKjETvAtzc,6478
|
|
873
|
+
wisent/core/parser_arguments/generate_pairs_from_task_parser.py,sha256=wEZLIXocqgLduufNMYTNm4YcQwerq53zp5YseCqiCbY,924
|
|
874
|
+
wisent/core/parser_arguments/generate_pairs_parser.py,sha256=nDTQ63bvjl-GcSrMdtcR_Pdh1dzbqtRnGaSCFhPO-4Q,1918
|
|
875
|
+
wisent/core/parser_arguments/generate_responses_parser.py,sha256=LnjIx7gOdysEMGwUIHFHlaEOE306qr4rXto9gzBSjkg,1345
|
|
876
|
+
wisent/core/parser_arguments/generate_vector_from_synthetic_parser.py,sha256=Dje2Xuq-Xmd3HVYzDCluifqJfW9jTZ5ccleUyS8ZpPk,4209
|
|
877
|
+
wisent/core/parser_arguments/generate_vector_from_task_parser.py,sha256=nhRA8qlGmT_S7i0M9wLmj0KpBKlxgKkNs4pAw5JN14w,4246
|
|
878
|
+
wisent/core/parser_arguments/generate_vector_parser.py,sha256=bf2PHAdvNZetXaSe3sKarvK6McIODPdmRSNRQN_rHto,3622
|
|
879
|
+
wisent/core/parser_arguments/get_activations_parser.py,sha256=pCBET7CsFz-8WFE3mb677o4W-vVa09wz-3YDMx-8ukA,2445
|
|
880
|
+
wisent/core/parser_arguments/inference_config_parser.py,sha256=-TAcjy8yTzDKWFc42JCmVtGq1MowQI-IauRNFNj_se0,1853
|
|
881
|
+
wisent/core/parser_arguments/main_parser.py,sha256=ddwMq6cmEN5a1qJ_xAm6T3b0z68bh-BvL9-RByaaw9w,11932
|
|
882
|
+
wisent/core/parser_arguments/model_config_parser.py,sha256=RlcSyQkj0uDlmJJE8sMq_CjvJ1VdCcjf8_BqspsSHd0,3210
|
|
883
|
+
wisent/core/parser_arguments/modify_weights_parser.py,sha256=077Gz6GLk1aR2k6rboU4tcVBgxHaTB01sSRx7ZqlzPU,9336
|
|
884
|
+
wisent/core/parser_arguments/monitor_parser.py,sha256=qo3vyyVmdZBkhGuhlHqbEzTiIKcdFIvG3IIuc-DIRQE,1234
|
|
885
|
+
wisent/core/parser_arguments/multi_steer_parser.py,sha256=kt3KBOQMELwullEK8QixbKbsF5yX0esSmSYDT6rN120,2138
|
|
886
|
+
wisent/core/parser_arguments/nonsense_parser.py,sha256=ALxU27Hn7ewB8IVe7bMS78MqukU4riouy492nlM4BUM,1096
|
|
887
|
+
wisent/core/parser_arguments/optimization_cache_parser.py,sha256=R54XzVY4GUgpXu8CIFIpmCdenDK6DFf99ZmWC8NOSVU,2999
|
|
888
|
+
wisent/core/parser_arguments/optimize_classification_parser.py,sha256=tzHAvOPIbXP8risT6eMICpsRfO6RETwP4s8qIAe60ms,4170
|
|
889
|
+
wisent/core/parser_arguments/optimize_parser.py,sha256=HTuhFspCAZkNcYhP69-JmaT91CRm_wV9IaF-BFvg_I8,4603
|
|
890
|
+
wisent/core/parser_arguments/optimize_sample_size_parser.py,sha256=OJEY2y89-9pbA7FnTgDvxoa3SXvogL5yM2QqStbpBCA,2836
|
|
891
|
+
wisent/core/parser_arguments/optimize_steering_parser.py,sha256=dESy4Rm_whCszH3LGF-VK7wBObmtEDcODO3EoCOnWe4,24581
|
|
892
|
+
wisent/core/parser_arguments/optimize_weights_parser.py,sha256=Vfo_a7W1H4tQ1_vSUeoo3SjEg07yanQCvvg1gQvk3Rs,13685
|
|
893
|
+
wisent/core/parser_arguments/synthetic_parser.py,sha256=99fGg_k9vQl80COuVZw5LOYLRQ0sDlZ4ozwsx0T--Qs,5135
|
|
894
|
+
wisent/core/parser_arguments/tasks_parser.py,sha256=EUoRxKy3I5ekv7FGORAGLMeECBg2QL_fIv8nurI9jsQ,22992
|
|
895
|
+
wisent/core/parser_arguments/train_unified_goodness_parser.py,sha256=5BICCLDk6UrUjDXiwsOIYJYx9ubPwwPhC70gAP45Rq0,4962
|
|
896
|
+
wisent/core/parser_arguments/utils.py,sha256=XTtssIt05HlAyO6kMPZvu4l_UIZSMDaV-gl50NO7qOw,3543
|
|
897
|
+
wisent/core/prompts/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
898
|
+
wisent/core/prompts/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
899
|
+
wisent/core/prompts/core/atom.py,sha256=yb3uF7a8Vu7QL_UssriwHlIiqDef43NmppFg7Yj_JTg,1519
|
|
900
|
+
wisent/core/prompts/core/prompt_formater.py,sha256=u9j4gvHnB7ubbawW9Sueaxf0RdEkYPS2I6NtGvqiCKE,5382
|
|
901
|
+
wisent/core/prompts/prompt_stratiegies/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
902
|
+
wisent/core/prompts/prompt_stratiegies/direct_completion.py,sha256=a4XDl1NHK1hYsAlKgVV3pMBwVc6plTuBnmjup6rU7oY,834
|
|
903
|
+
wisent/core/prompts/prompt_stratiegies/instruction_following.py,sha256=T9kYsP96LwkfO94lUqZTTB3NFxJhFPZlS80Sah4bDVc,846
|
|
904
|
+
wisent/core/prompts/prompt_stratiegies/multiple_choice.py,sha256=L4dOtocXte91Mn2zUhldlwPHbTeNDlfi3m00_JE4KIc,905
|
|
905
|
+
wisent/core/prompts/prompt_stratiegies/role_playing.py,sha256=Bl_4A-OJgepDB33iRLB-fTdCLmy4QvfGKJ7kIq1ApUE,1101
|
|
906
|
+
wisent/core/steering_methods/__init__.py,sha256=ULxu5kHL5jCyrGoPAUiOYtuxU_iTYYY7e0frABdflfk,1364
|
|
907
|
+
wisent/core/steering_methods/preflight.py,sha256=26Pxx7b3TBRSHfnKF2NWhYN2mjTCnbnd3k7advIqjfU,12009
|
|
908
|
+
wisent/core/steering_methods/registry.py,sha256=kD76e_gtK-LwdPO6qz1Ai-dH_jYSobfOVSY5qzaMeUg,22482
|
|
909
|
+
wisent/core/steering_methods/rotator.py,sha256=Lt_IW2I1u9h79KLyiTtLIocGhdW0DxcQylMyqC8gClo,4053
|
|
910
|
+
wisent/core/steering_methods/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
911
|
+
wisent/core/steering_methods/core/atoms.py,sha256=Dtzg1t6nVAOJYRvQLptQchyjhiSx8mEgv2ImtXNCT_s,5485
|
|
912
|
+
wisent/core/steering_methods/methods/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
913
|
+
wisent/core/steering_methods/methods/caa.py,sha256=xcbjey1-hiPHd8rpoKA2T00_OD9rtpzRaKeajGbbFxI,1754
|
|
914
|
+
wisent/core/steering_methods/methods/prism.py,sha256=Dl-6jR9nz9m63bpGpUfIvfaNJSIpsuKKJB5zPL1Xk9A,24103
|
|
915
|
+
wisent/core/steering_methods/methods/pulse.py,sha256=5lChkLMeJJg49Ibd8hCmciPLSwrQjkiuPoTaXVplV_g,24277
|
|
916
|
+
wisent/core/steering_methods/methods/titan.py,sha256=lQ_wjQxtLZvle0aa3YyLN1iy4ExLdWM4WaBWGBIalcQ,39041
|
|
917
|
+
wisent/core/synthetic/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
918
|
+
wisent/core/synthetic/cleaners/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
919
|
+
wisent/core/synthetic/cleaners/deduper_cleaner.py,sha256=XI19V8IpX5lGWasW43Qa2peEEcxsrAfmB_W9uKzw11U,1571
|
|
920
|
+
wisent/core/synthetic/cleaners/pairs_cleaner.py,sha256=t-xKYMfLsASk06EKfRM6rWYhEXv9cV7kjSaPhgLosG8,3860
|
|
921
|
+
wisent/core/synthetic/cleaners/refusaler_cleaner.py,sha256=Cd7DPbN1icgr9BQzWykwG3MiAtUB5XwnY777RQrW0Xk,5370
|
|
922
|
+
wisent/core/synthetic/cleaners/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
923
|
+
wisent/core/synthetic/cleaners/core/atoms.py,sha256=XUrgDhrPs77gCWq9y4VSDaC4WlylRrT1o9vrZjtpB7Y,1350
|
|
924
|
+
wisent/core/synthetic/cleaners/methods/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
925
|
+
wisent/core/synthetic/cleaners/methods/base_dedupers.py,sha256=eT_QNEwiuROIjmSG3fQXsEHmlwagBAdWk6uNeqxPEOY,11901
|
|
926
|
+
wisent/core/synthetic/cleaners/methods/base_refusalers.py,sha256=3l-w_kaFY9dfwOCjhmmA0S-_wmtKIkGxcZwnGOCLsm0,10329
|
|
927
|
+
wisent/core/synthetic/cleaners/methods/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
928
|
+
wisent/core/synthetic/cleaners/methods/core/atoms.py,sha256=kxD-CnS-u2GXAiezNYEQJ0asgxbMpmR4sXroSz5QjxE,1563
|
|
929
|
+
wisent/core/synthetic/db_instructions/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
930
|
+
wisent/core/synthetic/db_instructions/mini_dp.py,sha256=XElWJGkIG9DR5q4wcY11kJDJPKw6obRcfBce8PUwjyM,7047
|
|
931
|
+
wisent/core/synthetic/db_instructions/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
932
|
+
wisent/core/synthetic/db_instructions/core/atoms.py,sha256=jBazuD37hyBMIsbFNDjpwlGnxECAOuh4YT0fmpH8C6o,748
|
|
933
|
+
wisent/core/synthetic/generators/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
934
|
+
wisent/core/synthetic/generators/nonsense_generator.py,sha256=M92jIiXmXBows_FdF-cyo2JgvKQIWRmKEN4RacsgKdk,5790
|
|
935
|
+
wisent/core/synthetic/generators/pairs_generator.py,sha256=VgxeXwTe_Ctvms8b5oBZL-H5l2ck9rExfQf0FHxeIVE,12927
|
|
936
|
+
wisent/core/synthetic/generators/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
937
|
+
wisent/core/synthetic/generators/core/atoms.py,sha256=9wL0v38BCqn3y9LtoRkQsK_X3egjdYcPmFXH0mgFSWg,2290
|
|
938
|
+
wisent/core/synthetic/generators/diversities/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
939
|
+
wisent/core/synthetic/generators/diversities/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
940
|
+
wisent/core/synthetic/generators/diversities/core/core.py,sha256=TjSj5T7NE5kRH-ABcFqb1Hz_j3Z6F_TcV-95uHD5Xw8,2201
|
|
941
|
+
wisent/core/synthetic/generators/diversities/methods/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
942
|
+
wisent/core/synthetic/generators/diversities/methods/fast_diversity.py,sha256=uAsNKq_YGibLK53srbywcNS0-pdhmp9_yuocSRDp_mg,8556
|
|
943
|
+
wisent/core/tasks/__init__.py,sha256=ut6pYq-FESYGiL90Re_L-vB9-8tSxCcDVr-grVf1_t0,8104
|
|
944
|
+
wisent/core/tasks/aime_task.py,sha256=iRhEaZkk3ixQ_uOn_Cq3IE9ycwCZlKWmZiYgn352qNw,5477
|
|
945
|
+
wisent/core/tasks/file_task.py,sha256=355yZoecMepfv5UIlKSH5bVBIaHFFl5DSbWpWqQJrq4,7207
|
|
946
|
+
wisent/core/tasks/hle_task.py,sha256=YBSvAA4WnM4OXdsAz6ryh6fL41vPwVHrMHF5OJeLkE4,7043
|
|
947
|
+
wisent/core/tasks/hmmt_task.py,sha256=I0UlzbzJp2ElvhYpKbd4QwZcoMh2yX5xvoZb8RAlhC0,4802
|
|
948
|
+
wisent/core/tasks/livecodebench_task.py,sha256=laQH6W2HWNCL17MtYR_tcEYKiFif_L3jA3ZaVElHAe0,3879
|
|
949
|
+
wisent/core/tasks/livemathbench_task.py,sha256=mzC_Y-TZziXFL_AfJILCWQM5SGF7CPQQPaDh_cneVss,6462
|
|
950
|
+
wisent/core/tasks/lm_eval_task.py,sha256=KLTYT_stwgcfBccuaUewzUbiptOgKHzKAyGv5jY1eWk,19291
|
|
951
|
+
wisent/core/tasks/math500_task.py,sha256=e_qpzG_ppuZxoq3Vi40n_9-vOCf4lRAWO4zIiyUezgU,3124
|
|
952
|
+
wisent/core/tasks/polymath_task.py,sha256=lp9ALtKuyKZAkC-FXdGFBQhjTtZs_CEm7EMM6BEyCJw,5993
|
|
953
|
+
wisent/core/tasks/supergpqa_task.py,sha256=_axLCD7-ZoigxTfaYwVrzyHuo5woeJxKaBpWSYJZ1Tg,9069
|
|
954
|
+
wisent/core/tracking/__init__.py,sha256=rda2gaV_tFP6CChzx8TBQUrh_PObqvX35-BHJCxi1Q8,1150
|
|
955
|
+
wisent/core/tracking/latency.py,sha256=uKibYRodOurQoCh5MqfJwo_k5Ho4H5dFQ0p_nddHGOg,21909
|
|
956
|
+
wisent/core/tracking/memory.py,sha256=Fwg92c_M997uToNLIZvFlHbp3F5TTrUZsA_X25J-QfM,13201
|
|
957
|
+
wisent/core/trainers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
958
|
+
wisent/core/trainers/steering_trainer.py,sha256=19GNV7QX9mBl5awPwXANK6haW-Xf4vCyYKgdAjvvH0o,15759
|
|
959
|
+
wisent/core/trainers/core/__init__.py,sha256=zkv6f6Auk4dbJ5PTQP96gP2TUM55W3tc1flHLWBvI4M,1246
|
|
960
|
+
wisent/core/trainers/core/atoms.py,sha256=ycWk0G-7EIAEOnQL-o5_V5B8KTQ7CQUilGF4ibjighM,1536
|
|
961
|
+
wisent/core/utils/__init__.py,sha256=ee4tMv5W_qxtTwRS-n8ZCl69CgBUqPGjugrkomOhzhE,1419
|
|
962
|
+
wisent/core/utils/base_rotator.py,sha256=5seiluDYXeOARs7Z5RIHlfW__9nUbxir6KGCs0wUHwI,10018
|
|
963
|
+
wisent/core/utils/dataset_splits.py,sha256=1G8TTiNFbjLcf17LjVh9xxlAqC1oyCMnOTQckHxPN6E,5937
|
|
964
|
+
wisent/core/utils/device.py,sha256=zqen3KLYV8j5qUyryCTv5lxTW4Ygn8dzQSt_pgUW_qs,9419
|
|
965
|
+
wisent/core/weight_modification/__init__.py,sha256=UrQh_lVfQIAddYQDOkmVWY3BP4X6uuvltHY6KHPDu0g,4329
|
|
966
|
+
wisent/core/weight_modification/additive.py,sha256=USpIfiz2K486qisbs8nubhyfaGgMGzT4CJnGjAB17sU,11375
|
|
967
|
+
wisent/core/weight_modification/directional.py,sha256=GLcbn7IRQp2uqgl1t5uRA0IE9QukfyiZGNxt-YVdK4s,49253
|
|
968
|
+
wisent/core/weight_modification/export.py,sha256=zikHU9dGDswa61ZPd_xo61kAdNt46O1VOQm3HOSYfNY,10266
|
|
969
|
+
wisent/core/weight_modification/multi_direction.py,sha256=tpnEG54GHIzWmytt3ajUmwTbl2_ClpZS7AebJHHrkfA,12858
|
|
970
|
+
wisent/core/weight_modification/utils.py,sha256=h6QM5FrpvOu29_6YZNvs5bwBuJBvVgtBo18RQ_MW3_o,6907
|
|
971
|
+
wisent/examples/contrastive_pairs/humanization_human_vs_ai.json,sha256=7D6GpW-B99nXnhLj2q7WYZp5DWu8AMFzk3uyq7ambV0,247507
|
|
972
|
+
wisent/examples/scripts/benchmark_tags.json,sha256=JjZGitGdgbJPXxk6PXGDsEaQBvAKlwJOAcET4rucRWA,29279
|
|
973
|
+
wisent/examples/scripts/lm_eval_readme.json,sha256=X5bYjQrAUw031rjjFlrTlxxnllWGBSlqywF5kjh4mOw,89869
|
|
974
|
+
wisent/examples/scripts/1/test_basqueglue_evaluation.json,sha256=4dmo2Mu7zZ_mFhrn1EWn5grPkOhNI4XjKUYmhCoICl0,1898
|
|
975
|
+
wisent/examples/scripts/1/test_basqueglue_pairs.json,sha256=yCLswgp9q3SDzhso-pB5jbIlTHzqS_GEBtwCQgVwlXo,650
|
|
976
|
+
wisent/examples/scripts/1/test_bec2016eu_evaluation.json,sha256=rVRPrb-pot0bCu1p69ks2ltQiW_grUtRvxZonYupUY4,1897
|
|
977
|
+
wisent/examples/scripts/1/test_bec2016eu_pairs.json,sha256=yCLswgp9q3SDzhso-pB5jbIlTHzqS_GEBtwCQgVwlXo,650
|
|
978
|
+
wisent/examples/scripts/1/test_belebele_evaluation.json,sha256=xxQkvL-kBvLKhhsNsSCEni7Dw7BFoEaUMY69uJO99w0,2214
|
|
979
|
+
wisent/examples/scripts/1/test_belebele_pairs.json,sha256=h9RypNbw7geXsg7fWcBUwgJB5bS5gy0eDPGDizbouPQ,944
|
|
980
|
+
wisent/examples/scripts/1/test_benchmarks_evaluation.json,sha256=MAOCk2IdwkhJ_uDtTSNl1CrO2HxhXLwNSMTAUgZZ9Lc,1812
|
|
981
|
+
wisent/examples/scripts/1/test_benchmarks_pairs.json,sha256=KlTcdNMmXAlwArOmV3qLdfjTTSygO8Elu_eZWK3k-5A,1086
|
|
982
|
+
wisent/examples/scripts/1/test_bertaqa_evaluation.json,sha256=kto40P6Jo5UuPWvxjjLvTr_bN7MRSbATwIoWHff4BbY,1604
|
|
983
|
+
wisent/examples/scripts/1/test_bertaqa_pairs.json,sha256=7OtQCoFqK1_29r4zL8FSsF2c91KUKN4FYBeomdwfPL4,391
|
|
984
|
+
wisent/examples/scripts/1/test_bhtc_v2_evaluation.json,sha256=izOv-P-hWlmm-d27yPmscDSVLZ9UzLnX3E5nztZK10Q,904
|
|
985
|
+
wisent/examples/scripts/1/test_bhtc_v2_pairs.json,sha256=KZRti9y-wFKygQkcTM7LuimkUhcHd4qpgEGwCgPuw_0,427
|
|
986
|
+
wisent/examples/scripts/1/test_boolq-seq2seq_evaluation.json,sha256=T4qMKma5AtYicG6xFccx03GLYnLAyA-Wa-qmvACbIc4,872
|
|
987
|
+
wisent/examples/scripts/1/test_boolq-seq2seq_pairs.json,sha256=YZiJJSLsFTIRMfctwiXSEStI3YaznaWTbF58-eGE8OQ,1583
|
|
988
|
+
wisent/examples/scripts/1/test_cabreu_evaluation.json,sha256=mR963WuUoU0qoC4o_-3ZvLDPfwiN_kkv5QxU4WQcSj0,7307
|
|
989
|
+
wisent/examples/scripts/1/test_cabreu_pairs.json,sha256=RJT13aObmZiWP5f6dptqoRVM8_sznJxY8QEHH_iVeBY,6895
|
|
990
|
+
wisent/examples/scripts/1/test_careqa_en_evaluation.json,sha256=cngnpcX1UOUEyQo9DNx519jPjtiuN7uC946sish4AB4,1675
|
|
991
|
+
wisent/examples/scripts/1/test_careqa_en_pairs.json,sha256=Of0zuMtBLEafqCEKR5ffqC-PghQCzykMCT9N3Rp4NwI,731
|
|
992
|
+
wisent/examples/scripts/1/test_careqa_evaluation.json,sha256=B8C75kicOK3quYBZCDhFkh-LHJV7hkINoI_6NwGu-g4,1672
|
|
993
|
+
wisent/examples/scripts/1/test_careqa_pairs.json,sha256=Of0zuMtBLEafqCEKR5ffqC-PghQCzykMCT9N3Rp4NwI,731
|
|
994
|
+
wisent/examples/scripts/1/test_catalanqa_evaluation.json,sha256=bomYcxWdkdabZ7ptvrPWpugAJcVNpSbKtHlfW-JXuLs,918
|
|
995
|
+
wisent/examples/scripts/1/test_catalanqa_pairs.json,sha256=9zue9ntmKjFwKJetuLz3BX0HpBCL72000sjWT4IdkjA,952
|
|
996
|
+
wisent/examples/scripts/1/test_catcola_evaluation.json,sha256=w2zfhB6YQ7l8szinoIni_zi8ncbIGiUuIolXa9JDrDA,918
|
|
997
|
+
wisent/examples/scripts/1/test_catcola_pairs.json,sha256=DES1F-Bavgy_-7fl420bSb32fuSa-JQBjBwKuv1hMmE,240
|
|
998
|
+
wisent/examples/scripts/1/test_chartqa_evaluation.json,sha256=V9Yr1qLmw3BFtnTSet3DUXaIVZAgYq431mP4doqXB1g,819
|
|
999
|
+
wisent/examples/scripts/1/test_chartqa_pairs.json,sha256=bficR5iHawdvC465Gu2-26200Gd-4nT_MRn9Q57JVtY,163
|
|
1000
|
+
wisent/examples/scripts/1/test_claim_stance_topic_evaluation.json,sha256=r43kMchm9NeoBqvrwRGSmdf_RQ8DkR8k-CiPnpu_COY,917
|
|
1001
|
+
wisent/examples/scripts/1/test_claim_stance_topic_pairs.json,sha256=Q4DP2sKyGgPqnerSfkLpmjGeYjUuEqdkaHLbs6GfADM,1910
|
|
1002
|
+
wisent/examples/scripts/1/test_cnn_dailymail_evaluation.json,sha256=2xdFLCuLrI7Jhqu04QyhSPSjXPrwjKbPGhMV4kSjEl8,1710
|
|
1003
|
+
wisent/examples/scripts/1/test_cnn_dailymail_pairs.json,sha256=S1c_5HNR7YMIRy7z0fjMsRlCd5KhbPpmYXdgbl1WUIs,4727
|
|
1004
|
+
wisent/examples/scripts/1/test_cocoteros_es_evaluation.json,sha256=0EyWwlC_uDBECFVMxilzkx3fYwCrXDw-15nQjFChcS4,1181
|
|
1005
|
+
wisent/examples/scripts/1/test_cocoteros_es_pairs.json,sha256=ZzoRs2NiHtKditYWvCrR-udBw9ROFfEqKs1ZUrTp2r8,385
|
|
1006
|
+
wisent/examples/scripts/1/test_coedit_gec_evaluation.json,sha256=oNucJ-YJW0OrWf8Jmlbf6ie3lCvM_HcYY1nddcvpIN4,1762
|
|
1007
|
+
wisent/examples/scripts/1/test_coedit_gec_pairs.json,sha256=80t6Cfu9be8TpPNaF0WOsRME6YbfLVyIBJyLGEuqM9o,687
|
|
1008
|
+
wisent/examples/scripts/1/test_cola_evaluation.json,sha256=4TOycnPImbfXliX05dKc8CUTW2MR8cKihmH0DRKR7OY,915
|
|
1009
|
+
wisent/examples/scripts/1/test_cola_pairs.json,sha256=8a5Dd84Qbeuy2a3CVpNfDwX4PGhYuw60j7YsM_5YwYE,249
|
|
1010
|
+
wisent/examples/scripts/1/test_coqcat_evaluation.json,sha256=i5pKT9mUk6iS2ubiq_mkW4utZIP15oVFqCK76a-adpE,978
|
|
1011
|
+
wisent/examples/scripts/1/test_coqcat_pairs.json,sha256=GWzqrA3XI7CK6VElef36IanFOOS7uJU-uKJSQ_8lQgg,1414
|
|
1012
|
+
wisent/examples/scripts/1/test_dbpedia_14_evaluation.json,sha256=9IKJ7skQ9IChdGin5YH2C8FD4dRZwb6zYG4IypiZ4eA,889
|
|
1013
|
+
wisent/examples/scripts/1/test_dbpedia_14_pairs.json,sha256=5DCuHnpY4v9-JNSoQwOh-UG9bFBNF2CJi9RBqVTIyBc,443
|
|
1014
|
+
wisent/examples/scripts/1/test_epec_koref_bin_evaluation.json,sha256=uptYxyqdNCvVAMXKVBzOe7MD-iagHcaPDDHzBPkeIJU,875
|
|
1015
|
+
wisent/examples/scripts/1/test_epec_koref_bin_pairs.json,sha256=_D4-SAq-KYq4Dh7Sz4Je_grkJs2A16XhYRcjWHMDBfM,395
|
|
1016
|
+
wisent/examples/scripts/1/test_ethos_binary_evaluation.json,sha256=XeHT77YVOvGVUtIqu73B6FuxkxsdtWh8-C_U-5dYhsQ,939
|
|
1017
|
+
wisent/examples/scripts/1/test_ethos_binary_pairs.json,sha256=mfqI7rEbpZlnyMCufGG5QcCyq4iOSEJHZiKEbY7Mm7M,449
|
|
1018
|
+
wisent/examples/scripts/2/test_afrimgsm_direct_amh_evaluation.json,sha256=bDPS5GxmsmKRb6sAxErR4nogXmgSOxzIP_BOzzf9EQ8,1271
|
|
1019
|
+
wisent/examples/scripts/2/test_afrimgsm_direct_amh_pairs.json,sha256=xtGe0bshz2iLAbIuI12IUeOI-6LpugfhSbXUWi8dYfQ,1000
|
|
1020
|
+
wisent/examples/scripts/2/test_afrimmlu_direct_amh_evaluation.json,sha256=eM0Rp54yfE6xswWrx4CCXRHj7Wo7ZJA5Q8i0V0lOz-0,1190
|
|
1021
|
+
wisent/examples/scripts/2/test_afrimmlu_direct_amh_pairs.json,sha256=DLJ7qHdvIdpzAFTDDlGqdFqYQLmdjHsVbq5PgjnrrjU,669
|
|
1022
|
+
wisent/examples/scripts/2/test_afrixnli_en_direct_amh_evaluation.json,sha256=qNsSpTnnRG8yVEfYeYV4sixnxg1XrSzrj0I0Rkkesag,1080
|
|
1023
|
+
wisent/examples/scripts/2/test_afrixnli_en_direct_amh_pairs.json,sha256=BKbzJUR8HjFLi7IGYudSXIx94pGh9x4AvO-ZkOndsRc,386
|
|
1024
|
+
wisent/examples/scripts/2/test_arc_ar_evaluation.json,sha256=siu_sO29iLp7Dz_K475pps8TPYTwBfmxr7bPzfG6zXw,2358
|
|
1025
|
+
wisent/examples/scripts/2/test_arc_ar_pairs.json,sha256=dN5eXC6lL2WF3cpuZKOJEtTXM7P7EpFHvSkheFrMw8M,1188
|
|
1026
|
+
wisent/examples/scripts/2/test_atis_evaluation.json,sha256=pv1lWUOHOPfQUxhu8S5dvfoypxSBTPxaknMv2D9baNc,1365
|
|
1027
|
+
wisent/examples/scripts/2/test_atis_pairs.json,sha256=EieW893xtNjVBpJrJxJuTy1OvHnnZzJKPctnH1TevVE,1916
|
|
1028
|
+
wisent/examples/scripts/2/test_babi_evaluation.json,sha256=2O0czob2UoLANSIBO0BEAVNe4-gTOPdPH3sFcDA3u0k,899
|
|
1029
|
+
wisent/examples/scripts/2/test_babi_pairs.json,sha256=SjxpSXlWLOHx26XEozN6rTiYfwrkBG8rqies72sekLw,222
|
|
1030
|
+
wisent/examples/scripts/2/test_babilong_evaluation.json,sha256=sBiv35EeORDv4TOQRx3yd_SoB99LblbsTznrFbeEWUk,1454
|
|
1031
|
+
wisent/examples/scripts/2/test_babilong_pairs.json,sha256=OOWUdyqtMXUv2_WWIqTTEYX9okJXGVzIgnSYgAGljuc,41866
|
|
1032
|
+
wisent/examples/scripts/2/test_bangla_mmlu_evaluation.json,sha256=OtR2JvARwR99QphQDprHiyv7wjLPK9583gR_rCM9RQU,2073
|
|
1033
|
+
wisent/examples/scripts/2/test_bangla_mmlu_pairs.json,sha256=GR9eXDwYdp0CBOyOZ9XF_yHv07KlDcpv8BpPF5wovXw,1340
|
|
1034
|
+
wisent/examples/scripts/2/test_basque-glue_pairs.json,sha256=yCLswgp9q3SDzhso-pB5jbIlTHzqS_GEBtwCQgVwlXo,650
|
|
1035
|
+
wisent/examples/scripts/results/benchmark_descriptions.json,sha256=zAYiDpmznc1xxFPGo1Np6cFLoeI-AnGpyKKmaHBYiKE,72438
|
|
1036
|
+
wisent/examples/scripts/results/benchmark_evaluation_methods.json,sha256=AL4c5mOLnUS4Fx-Qw6ZoeuQ-oRqNSjl0N9lBaGYAQAU,2058
|
|
1037
|
+
wisent/examples/scripts/results/benchmark_evaluator_mapping.json,sha256=al6hXKMSzM765PHeAOWcitHosNCxy4gMK9NIaA9QMyE,98824
|
|
1038
|
+
wisent/examples/scripts/results/benchmark_evaluator_mapping_updated.json,sha256=wwWp5RsEbII7wo17bAmqmwi9hZk5Vdez1Gt2GFSmMjg,1259468
|
|
1039
|
+
wisent/examples/scripts/results/benchmark_evaluators_clean.json,sha256=QFZxJRwzHz6TTVbtS-FJJVwa4M5N9vrI3qeZ8NqtGDw,29351
|
|
1040
|
+
wisent/examples/scripts/results/benchmark_methods_summary.json,sha256=xdH8RVEQOusqsPFxTisYaL19OdogjMXxPwtAPU8qZKw,7747
|
|
1041
|
+
wisent/examples/scripts/results/benchmark_pair_creation_methods.json,sha256=qGuknhpHHMVGrnuWyOcaukn7JM1olpBE-IulXZnms6k,2444
|
|
1042
|
+
wisent/examples/scripts/results/benchmark_pair_totals.json,sha256=4QxerUtba3olSFBdFyjaY5DLXuc5Ml_TWP7_Z0s5AqU,5745
|
|
1043
|
+
wisent/examples/scripts/results/benchmark_tags.json,sha256=X_ouWT0yAz5tXEcoe6mEK2z8FoMxl3hyFKQFa6DFJtA,18479
|
|
1044
|
+
wisent/examples/scripts/results/benchmark_test_summary_nov4.json,sha256=Alj610XpssUmA4TUIBA8XahaPkZhygpW9Qxxp31Kw50,1248
|
|
1045
|
+
wisent/examples/scripts/results/coding_benchmarks_test_code_status.json,sha256=fdK0WRug8ZIc0xUVhc_6Qzik92BF_BWoaoHbGEbfyXI,5550
|
|
1046
|
+
wisent/examples/scripts/results/failing_benchmarks.json,sha256=q1QOAIFU5oUvU3YEQvIluTIaRAnyQbm1fi7wsrJjw9g,36471
|
|
1047
|
+
wisent/examples/scripts/results/failing_benchmarks_list.json,sha256=yLjh4irOrwutzpF6OQGwC3lxLJRGpvfrZxsEMHjIIuE,1625
|
|
1048
|
+
wisent/examples/scripts/results/failing_benchmarks_test_results.json,sha256=ziouxsa8VO_f5AuVuJtWmTgz_v4Zuh9nIDHnqigaU4c,18149
|
|
1049
|
+
wisent/examples/scripts/results/missing_benchmark_tags.json,sha256=UoXptlWIEX3edEJbAiJ06Do8-H3uHOMgioO9tUZyr10,4706
|
|
1050
|
+
wisent/examples/scripts/results/test_20_newsgroups_evaluation.json,sha256=UJXxIGeFmb-qoQCL1V12X5csTiGMA7yMPvk4pNHCWWU,916
|
|
1051
|
+
wisent/examples/scripts/results/test_20_newsgroups_pairs.json,sha256=w7GNDUp6hf-FSGR8Z2djvnuYDKZZHynnnBZb-kg-Agg,1124
|
|
1052
|
+
wisent/examples/scripts/results/test_AraDICE_evaluation.json,sha256=2lniZTUZ-g5TQFjW-_daDD_Zl7HaFiqzpDzcGOXSPUY,2447
|
|
1053
|
+
wisent/examples/scripts/results/test_AraDICE_pairs.json,sha256=ZcTWPtrbEnTABLf1t8MRVuMW6UP4zdJ5FuC7BBEnwus,1076
|
|
1054
|
+
wisent/examples/scripts/results/test_ArabCulture_evaluation.json,sha256=dDToZso53F7E2NT-rWxgPtSqqrBoNWRC5q-4KsbDes4,5664
|
|
1055
|
+
wisent/examples/scripts/results/test_ArabCulture_pairs.json,sha256=0WKa_XqsgWS3NrPuzg_bwTFW3xcm4D7BTsGaiOSMMnw,2742
|
|
1056
|
+
wisent/examples/scripts/results/test_Tag_evaluation.json,sha256=diHsOMTBW4AVNg-GU6CaF6etSofZLEVdWa9cRPrN-s4,1073
|
|
1057
|
+
wisent/examples/scripts/results/test_Tag_pairs.json,sha256=anAOb_Saxi_l_RZ8cE78KMEwMQIn0dX3zsWN_ipsPrA,384
|
|
1058
|
+
wisent/examples/scripts/results/test_aclue_evaluation.json,sha256=dlh3Nch72S0a4qyf-YzaM4VxCR2-DaViH_EH5VJrZCw,4111
|
|
1059
|
+
wisent/examples/scripts/results/test_aclue_pairs.json,sha256=-YrdFc_L2OzAmtkM0BAus1qfhGJfYl7pYL3Gvssy-to,1859
|
|
1060
|
+
wisent/examples/scripts/results/test_acp_bench_evaluation.json,sha256=h5D1gkKgKVY9ieT4jwcYjnBGF_1VajJ7fiIvaGI2uhY,1589
|
|
1061
|
+
wisent/examples/scripts/results/test_acp_bench_hard_evaluation.json,sha256=9Vb7T7ChOs_TBsHS7piJ-x1scgJyJ7kwilMFTmIuiy8,2804
|
|
1062
|
+
wisent/examples/scripts/results/test_acp_bench_hard_pairs.json,sha256=ZT8S5GIgWwPyfNtf8jHEO1_SAApmVx5dt6TgLei_W0o,2544
|
|
1063
|
+
wisent/examples/scripts/results/test_acp_bench_pairs.json,sha256=D3L4hYxhS0J3JBsk2izEmXhxt5SSWbng76vsA6sfI-4,1652
|
|
1064
|
+
wisent/examples/scripts/results/test_advanced_ai_risk_evaluation.json,sha256=0rbRqeXUxYIIk-QxNva2WMYoFzpATSHgGnfJufACaUM,1604
|
|
1065
|
+
wisent/examples/scripts/results/test_advanced_ai_risk_pairs.json,sha256=i7Sg5QIV1_E8bb4jT6AXkNvvyBrNeW9PVl6YdxFJb5U,590
|
|
1066
|
+
wisent/examples/scripts/results/test_aexams_evaluation.json,sha256=fCdtf1V6lf-qviXWgsN-bnq9t7WeQ-3jvhAkxJsrMO4,2768
|
|
1067
|
+
wisent/examples/scripts/results/test_aexams_pairs.json,sha256=1evSD_zurT_bTUQkdqBYq-xt4KoPeFcUB5Z20-BACpI,1138
|
|
1068
|
+
wisent/examples/scripts/results/test_afrimgsm_direct_amh_evaluation.json,sha256=bDPS5GxmsmKRb6sAxErR4nogXmgSOxzIP_BOzzf9EQ8,1271
|
|
1069
|
+
wisent/examples/scripts/results/test_afrimgsm_direct_amh_pairs.json,sha256=xtGe0bshz2iLAbIuI12IUeOI-6LpugfhSbXUWi8dYfQ,1000
|
|
1070
|
+
wisent/examples/scripts/results/test_afrimmlu_direct_amh_evaluation.json,sha256=eM0Rp54yfE6xswWrx4CCXRHj7Wo7ZJA5Q8i0V0lOz-0,1190
|
|
1071
|
+
wisent/examples/scripts/results/test_afrimmlu_direct_amh_pairs.json,sha256=DLJ7qHdvIdpzAFTDDlGqdFqYQLmdjHsVbq5PgjnrrjU,669
|
|
1072
|
+
wisent/examples/scripts/results/test_afrixnli_en_direct_amh_evaluation.json,sha256=qNsSpTnnRG8yVEfYeYV4sixnxg1XrSzrj0I0Rkkesag,1080
|
|
1073
|
+
wisent/examples/scripts/results/test_afrixnli_en_direct_amh_pairs.json,sha256=BKbzJUR8HjFLi7IGYudSXIx94pGh9x4AvO-ZkOndsRc,386
|
|
1074
|
+
wisent/examples/scripts/results/test_ag_news_evaluation.json,sha256=9oF_oOxv8M4X4pszreBA3V6wvvdO2mBxuqyMzH4flbI,894
|
|
1075
|
+
wisent/examples/scripts/results/test_ag_news_pairs.json,sha256=gjrHIcGIAjlYhn4a2ebIwER6BxA_7Y7Xp9XczzOFfTs,479
|
|
1076
|
+
wisent/examples/scripts/results/test_agieval_evaluation.json,sha256=S2pRGt0RfH5roC6wRVyYjZcpa3ZG4Om9dSTLEiVd784,1628
|
|
1077
|
+
wisent/examples/scripts/results/test_agieval_pairs.json,sha256=us7zfgCQMGb4ZGSxdSLgR89P-25_Uop_RgdYQt1E0Ls,664
|
|
1078
|
+
wisent/examples/scripts/results/test_aime2024_evaluation.json,sha256=iG-4CFZRQZI6N2PM-_DQAwht9D10IOaOkBf3ZG5YHtY,867
|
|
1079
|
+
wisent/examples/scripts/results/test_aime2024_pairs.json,sha256=16Oxx8g2g21iKaUADksGSPKVY00i7I7C3PcczBoWbcA,221
|
|
1080
|
+
wisent/examples/scripts/results/test_aime2025_evaluation.json,sha256=zeuyGWCDGicdpXctuf13NhgG34O9CIMnkqpBg7UvoUM,867
|
|
1081
|
+
wisent/examples/scripts/results/test_aime2025_pairs.json,sha256=16Oxx8g2g21iKaUADksGSPKVY00i7I7C3PcczBoWbcA,221
|
|
1082
|
+
wisent/examples/scripts/results/test_aime_evaluation.json,sha256=lPUpLNXGLc2mIv7J7M-rDEXeiy5PR8vNrq1l_I88KeU,863
|
|
1083
|
+
wisent/examples/scripts/results/test_aime_pairs.json,sha256=16Oxx8g2g21iKaUADksGSPKVY00i7I7C3PcczBoWbcA,221
|
|
1084
|
+
wisent/examples/scripts/results/test_anagrams1_evaluation.json,sha256=9WCkRaLkJdLe-l3B3EkGM962EK_d_IXA99EyoWpcPAk,798
|
|
1085
|
+
wisent/examples/scripts/results/test_anagrams1_pairs.json,sha256=4JPuIVLPonmZp1aryksfHTy7yvq5TilpRjgBlfsHGD4,124
|
|
1086
|
+
wisent/examples/scripts/results/test_anagrams2_evaluation.json,sha256=VpI6UcF9XjvcL_8FIoJ-BpwiO9MNhzyRQRQcdGKot4U,828
|
|
1087
|
+
wisent/examples/scripts/results/test_anagrams2_pairs.json,sha256=KxxhxCgmBPi6zob3r__7kMGrI2mQtXw1Y-cq3qYQltw,137
|
|
1088
|
+
wisent/examples/scripts/results/test_anli_evaluation.json,sha256=9abyL6Jh4J9wM8dW0g-FvtqoNHqTLFqv0DMPHBa4790,885
|
|
1089
|
+
wisent/examples/scripts/results/test_anli_pairs.json,sha256=3CaPgA3zSy6PaHucuK4paSX411Gx18N2LOZftJ_OpPE,525
|
|
1090
|
+
wisent/examples/scripts/results/test_apps_evaluation.json,sha256=yuOyofvIOBv7zCu328X5rRcskb7YrJHezwvNWnDNqUM,9529
|
|
1091
|
+
wisent/examples/scripts/results/test_apps_pairs.json,sha256=qym_IuMoP3pbBSWplvQ0cMnAwl0z-0kcRmheUQUH5Ao,5746
|
|
1092
|
+
wisent/examples/scripts/results/test_arabic_exams_evaluation.json,sha256=F6IVmGHJsRGI1AhpG25N_zQdrWy5FoF8cT0GDobtNhk,1123
|
|
1093
|
+
wisent/examples/scripts/results/test_arabic_exams_pairs.json,sha256=9mUDUuSksBUB1L1DzAyYlcIL8OSxY552i5v2L59WRFI,446
|
|
1094
|
+
wisent/examples/scripts/results/test_arabic_leaderboard_complete_evaluation.json,sha256=04MoEweDm8ghjHplVEl4F7v8yib_ATO1XhJ3aBcFftw,2266
|
|
1095
|
+
wisent/examples/scripts/results/test_arabic_leaderboard_complete_pairs.json,sha256=epwwxpn8d7LRlTUZg_aAa15hCxEx3_enKgRlMUeNxkI,989
|
|
1096
|
+
wisent/examples/scripts/results/test_arabic_leaderboard_light_evaluation.json,sha256=ZFFMu6-QFGXl32-Sjnbo8Eb4pckd9U9QfBOqhQgKhbc,2036
|
|
1097
|
+
wisent/examples/scripts/results/test_arabic_leaderboard_light_pairs.json,sha256=1kQfNNeVRMPyBTinXK9ka1kWSHyNCR6H9QSddSNRP1o,762
|
|
1098
|
+
wisent/examples/scripts/results/test_arabicmmlu_evaluation.json,sha256=hcxHHt2_dzZHr5YNjQikEl8mg8CDzAAugARjgcNSP3Q,4028
|
|
1099
|
+
wisent/examples/scripts/results/test_arabicmmlu_pairs.json,sha256=qJzEIlg4SDZsL5QCDmfSqxnrVqwa7RIwFjKFZ3Vm4Cw,2092
|
|
1100
|
+
wisent/examples/scripts/results/test_arc_ar_evaluation.json,sha256=siu_sO29iLp7Dz_K475pps8TPYTwBfmxr7bPzfG6zXw,2358
|
|
1101
|
+
wisent/examples/scripts/results/test_arc_ar_pairs.json,sha256=dN5eXC6lL2WF3cpuZKOJEtTXM7P7EpFHvSkheFrMw8M,1188
|
|
1102
|
+
wisent/examples/scripts/results/test_arc_challenge_evaluation.json,sha256=gAlZ2puwR8LqvMC7nFUSzdBJke5K-wx1hD4XO8YTwRo,1070
|
|
1103
|
+
wisent/examples/scripts/results/test_arc_challenge_pairs.json,sha256=i_ZNQATvJAvvUh3IyoEZaVz2PPK3e9qakj96jTkX4jw,415
|
|
1104
|
+
wisent/examples/scripts/results/test_arc_easy_evaluation.json,sha256=u6Fz5JyIp3XNxOqgyCuK322JCSKKpsOAeyV7NznaqDQ,938
|
|
1105
|
+
wisent/examples/scripts/results/test_arc_easy_pairs.json,sha256=ILijbNk0MRN8QVigMTJj20iWI-cT0htqemvDRbo8NSU,225
|
|
1106
|
+
wisent/examples/scripts/results/test_argument_topic_evaluation.json,sha256=n2rP_sHh-8uT4z49Vn6j_jiCrepM89sSPZCPPd_DdqA,959
|
|
1107
|
+
wisent/examples/scripts/results/test_argument_topic_pairs.json,sha256=K89ZeRQQPd5Y8GxfsVXFiDQl8rJ86FU9VhiDj-LNfnk,1630
|
|
1108
|
+
wisent/examples/scripts/results/test_arithmetic_evaluation.json,sha256=2DZcnt5HyiWLiGq1nf-IC036lb5Sz3iVE2jPk0-RpkQ,1514
|
|
1109
|
+
wisent/examples/scripts/results/test_arithmetic_pairs.json,sha256=c31tGjv2mG7qqbymz5DneDxlYO4wMbhrVUUQUDeRrw4,334
|
|
1110
|
+
wisent/examples/scripts/results/test_asdiv_evaluation.json,sha256=HAwj7V8cHtyM4j7GwWbasl6zXZepW0mYZNj0OWycq2k,859
|
|
1111
|
+
wisent/examples/scripts/results/test_asdiv_pairs.json,sha256=9uIoQJXdSnpn2HYFH_sYeUz78vpeRi7Nz8VS5Zrj7q8,221
|
|
1112
|
+
wisent/examples/scripts/results/test_assin_entailment_evaluation.json,sha256=ogZaWg1VDrmYs8xjubr_yGtwmjPmpXgsrtD3ewzxq04,2146
|
|
1113
|
+
wisent/examples/scripts/results/test_assin_entailment_pairs.json,sha256=0Qt0R3rGcSzEciIbX8fFJxZTsdtdnYTGXS2p55l9WKI,633
|
|
1114
|
+
wisent/examples/scripts/results/test_atis_evaluation.json,sha256=pv1lWUOHOPfQUxhu8S5dvfoypxSBTPxaknMv2D9baNc,1365
|
|
1115
|
+
wisent/examples/scripts/results/test_atis_pairs.json,sha256=EieW893xtNjVBpJrJxJuTy1OvHnnZzJKPctnH1TevVE,1916
|
|
1116
|
+
wisent/examples/scripts/results/test_babi_evaluation.json,sha256=2O0czob2UoLANSIBO0BEAVNe4-gTOPdPH3sFcDA3u0k,899
|
|
1117
|
+
wisent/examples/scripts/results/test_babi_pairs.json,sha256=SjxpSXlWLOHx26XEozN6rTiYfwrkBG8rqies72sekLw,222
|
|
1118
|
+
wisent/examples/scripts/results/test_babilong_evaluation.json,sha256=sBiv35EeORDv4TOQRx3yd_SoB99LblbsTznrFbeEWUk,1454
|
|
1119
|
+
wisent/examples/scripts/results/test_babilong_pairs.json,sha256=OOWUdyqtMXUv2_WWIqTTEYX9okJXGVzIgnSYgAGljuc,41866
|
|
1120
|
+
wisent/examples/scripts/results/test_bangla_mmlu_evaluation.json,sha256=OtR2JvARwR99QphQDprHiyv7wjLPK9583gR_rCM9RQU,2073
|
|
1121
|
+
wisent/examples/scripts/results/test_bangla_mmlu_pairs.json,sha256=GR9eXDwYdp0CBOyOZ9XF_yHv07KlDcpv8BpPF5wovXw,1340
|
|
1122
|
+
wisent/examples/scripts/results/test_banking77_evaluation.json,sha256=4qs2s_yVg7iYEp9u5O58tCAUHxg7qQqBkkyHovV89w4,982
|
|
1123
|
+
wisent/examples/scripts/results/test_banking77_pairs.json,sha256=26xg0AmI1Z4DPKwdvCZt2jwjVpOnMmMv3Rs6cyV4y0o,1991
|
|
1124
|
+
wisent/examples/scripts/results/test_basque-glue_evaluation.json,sha256=WRLIktFz9oQIcpHp0zd4HCStK9G-LzFX3t8LNao6zFU,2955
|
|
1125
|
+
wisent/examples/scripts/results/test_basque-glue_pairs.json,sha256=I5cwPeJcJXsaRrao0X_Q8iJCcCAzBPjNOaQxQIymgRA,1228
|
|
1126
|
+
wisent/examples/scripts/results/test_basque_bench_evaluation.json,sha256=qbmcyxku2K_K14Jh7pjT593oiRLX11c3pnqZNGUc_hc,2956
|
|
1127
|
+
wisent/examples/scripts/results/test_basque_bench_pairs.json,sha256=JzsmkN33b71HOLaZZAyBX9pdfgRZlDf_bIM-frz4a00,1228
|
|
1128
|
+
wisent/examples/scripts/results/test_basqueglue_evaluation.json,sha256=0KYY8wMMpHuvdtlsV9CmQjEW1xJ9-zOEnzEZIsSQzLg,2954
|
|
1129
|
+
wisent/examples/scripts/results/test_basqueglue_pairs.json,sha256=YzUEnHI38ZhDKsbBNOsnYy1OmjKTF3gZdgJSa5a-IAk,1228
|
|
1130
|
+
wisent/examples/scripts/results/test_bbh_evaluation.json,sha256=XcZ_04q_bB-0vgGoE0hiqB_nppKvnISTpcja_H0dvFo,2135
|
|
1131
|
+
wisent/examples/scripts/results/test_bbh_pairs.json,sha256=7KCwuccPW3lugOJWvUUlHpxhr4wZ3UzFf_wy_ssbC_A,1311
|
|
1132
|
+
wisent/examples/scripts/results/test_bbq_evaluation.json,sha256=-Lo6RQXt_XkMhWzwOzHBI39dbVlSRa7jspBuxSAXod4,954
|
|
1133
|
+
wisent/examples/scripts/results/test_bbq_pairs.json,sha256=26v1q-03Pb42IVkyfhgNLxSWh34AZ_atA9HGzo-EvyQ,331
|
|
1134
|
+
wisent/examples/scripts/results/test_bec2016eu_evaluation.json,sha256=DfqUQphUbPLSFnMRRo4yy07qSoQG7T79uRR8FhnMJ_A,2953
|
|
1135
|
+
wisent/examples/scripts/results/test_bec2016eu_pairs.json,sha256=7lNJz42y25-SMuIrhUlfJ7BHyek5x36r8aHcNhmSYSs,1228
|
|
1136
|
+
wisent/examples/scripts/results/test_belebele_evaluation.json,sha256=xxQkvL-kBvLKhhsNsSCEni7Dw7BFoEaUMY69uJO99w0,2214
|
|
1137
|
+
wisent/examples/scripts/results/test_belebele_pairs.json,sha256=h9RypNbw7geXsg7fWcBUwgJB5bS5gy0eDPGDizbouPQ,944
|
|
1138
|
+
wisent/examples/scripts/results/test_benchmarks_evaluation.json,sha256=MAOCk2IdwkhJ_uDtTSNl1CrO2HxhXLwNSMTAUgZZ9Lc,1812
|
|
1139
|
+
wisent/examples/scripts/results/test_benchmarks_pairs.json,sha256=KlTcdNMmXAlwArOmV3qLdfjTTSygO8Elu_eZWK3k-5A,1086
|
|
1140
|
+
wisent/examples/scripts/results/test_bertaqa_evaluation.json,sha256=kto40P6Jo5UuPWvxjjLvTr_bN7MRSbATwIoWHff4BbY,1604
|
|
1141
|
+
wisent/examples/scripts/results/test_bertaqa_pairs.json,sha256=7OtQCoFqK1_29r4zL8FSsF2c91KUKN4FYBeomdwfPL4,391
|
|
1142
|
+
wisent/examples/scripts/results/test_bhtc_v2_evaluation.json,sha256=izOv-P-hWlmm-d27yPmscDSVLZ9UzLnX3E5nztZK10Q,904
|
|
1143
|
+
wisent/examples/scripts/results/test_bhtc_v2_pairs.json,sha256=KZRti9y-wFKygQkcTM7LuimkUhcHd4qpgEGwCgPuw_0,427
|
|
1144
|
+
wisent/examples/scripts/results/test_bigbench_evaluation.json,sha256=JyskONgf5iIJu-_klbZ4xMHO9XSV0_Z8z_VwkVM7fjU,1664
|
|
1145
|
+
wisent/examples/scripts/results/test_bigbench_pairs.json,sha256=rrh2TwJtjMmU2HoIUCAWWgaJQ5qDCyhBz4d2XvVDWNU,1799
|
|
1146
|
+
wisent/examples/scripts/results/test_blimp_evaluation.json,sha256=k-t20pwFwWEamVh8UGzPgKQCrYEVUiFYLsxZ7VhW0nE,2251
|
|
1147
|
+
wisent/examples/scripts/results/test_blimp_pairs.json,sha256=yiDwzVlle_cyQ6Ro9jnjcRohtoFMsXTNHqivjIOv3ao,558
|
|
1148
|
+
wisent/examples/scripts/results/test_boolq-seq2seq_evaluation.json,sha256=T4qMKma5AtYicG6xFccx03GLYnLAyA-Wa-qmvACbIc4,872
|
|
1149
|
+
wisent/examples/scripts/results/test_boolq-seq2seq_pairs.json,sha256=YZiJJSLsFTIRMfctwiXSEStI3YaznaWTbF58-eGE8OQ,1583
|
|
1150
|
+
wisent/examples/scripts/results/test_boolq_evaluation.json,sha256=b4iXD4lHIrUZjntmXJSXQ-ONFpsdQ_nh7zLSiev-HlA,864
|
|
1151
|
+
wisent/examples/scripts/results/test_boolq_pairs.json,sha256=jK_jGZrMSaJVwiBYYmG9FxYaQhpn4-J9HH5lflr0lZ4,1554
|
|
1152
|
+
wisent/examples/scripts/results/test_c4_evaluation.json,sha256=WWCj0Hx4tbJ5cfQBpMRHY4gop2l4ofhLG7h0Q6M0I5M,1288
|
|
1153
|
+
wisent/examples/scripts/results/test_c4_pairs.json,sha256=WT3CZayRSj_RMC_Uoi057gj5WrsArkpQZx0nDS5t2xw,371
|
|
1154
|
+
wisent/examples/scripts/results/test_cabreu_evaluation.json,sha256=LOrsAwYzppKWpoClOyTxBKDcpIP8Gqf6-vxbkMo85ys,7307
|
|
1155
|
+
wisent/examples/scripts/results/test_cabreu_pairs.json,sha256=PF0Ml8b3oo2-KEHpmYOFa_aZirNeK-woqOAwtZ3f6OE,6895
|
|
1156
|
+
wisent/examples/scripts/results/test_careqa_evaluation.json,sha256=B8C75kicOK3quYBZCDhFkh-LHJV7hkINoI_6NwGu-g4,1672
|
|
1157
|
+
wisent/examples/scripts/results/test_careqa_pairs.json,sha256=Of0zuMtBLEafqCEKR5ffqC-PghQCzykMCT9N3Rp4NwI,731
|
|
1158
|
+
wisent/examples/scripts/results/test_catalan_bench_evaluation.json,sha256=sJWE9OcP2gFWTyaDJiim_O4O6BIY0XQmTUR46_Ve9lI,1829
|
|
1159
|
+
wisent/examples/scripts/results/test_catalan_bench_pairs.json,sha256=ybJuUti0sCeyZuGFSdsTCfkKtiNSCjksSHc1VeLczys,1883
|
|
1160
|
+
wisent/examples/scripts/results/test_catalanqa_evaluation.json,sha256=bomYcxWdkdabZ7ptvrPWpugAJcVNpSbKtHlfW-JXuLs,918
|
|
1161
|
+
wisent/examples/scripts/results/test_catalanqa_pairs.json,sha256=9zue9ntmKjFwKJetuLz3BX0HpBCL72000sjWT4IdkjA,952
|
|
1162
|
+
wisent/examples/scripts/results/test_catcola_evaluation.json,sha256=w2zfhB6YQ7l8szinoIni_zi8ncbIGiUuIolXa9JDrDA,918
|
|
1163
|
+
wisent/examples/scripts/results/test_catcola_pairs.json,sha256=DES1F-Bavgy_-7fl420bSb32fuSa-JQBjBwKuv1hMmE,240
|
|
1164
|
+
wisent/examples/scripts/results/test_cb_evaluation.json,sha256=4RdWDmJ3xDINLKxeyIf09kqcGWWcnenUs8O2qpnSjfQ,881
|
|
1165
|
+
wisent/examples/scripts/results/test_cb_pairs.json,sha256=QKViqRnSVeb-nQgJL1zRvNVZ2r_UcgXgX3yrePZx7Es,329
|
|
1166
|
+
wisent/examples/scripts/results/test_ceval_evaluation.json,sha256=wcjDYFj5B8R7UcHxLMLFQR-7w97eIk0nRVl3lf1tPuc,2283
|
|
1167
|
+
wisent/examples/scripts/results/test_ceval_pairs.json,sha256=dk4jGjqCE_HZLAWz6lIivYWRAt6rz0cao2GhhElzGVg,872
|
|
1168
|
+
wisent/examples/scripts/results/test_chain_of_thought_evaluation.json,sha256=NIsj71ylbuDJqB_dEAmcnGf2hwvFjb4UxgwHbYwTZrY,1582
|
|
1169
|
+
wisent/examples/scripts/results/test_chain_of_thought_pairs.json,sha256=Hj0R614SWLTerw8Z-TtuY_ezknF4xVmt2RD1m7beZek,553
|
|
1170
|
+
wisent/examples/scripts/results/test_chartqa_evaluation.json,sha256=DJacdQkemuDjwuFuqhXjznhWLWPhnea9C-f4_GYcCSE,819
|
|
1171
|
+
wisent/examples/scripts/results/test_chartqa_pairs.json,sha256=lVc_WfpfxUY6JWT1f8hcQQlOA0kRLq5BDRKD-sTrdHU,163
|
|
1172
|
+
wisent/examples/scripts/results/test_claim_stance_topic_evaluation.json,sha256=r43kMchm9NeoBqvrwRGSmdf_RQ8DkR8k-CiPnpu_COY,917
|
|
1173
|
+
wisent/examples/scripts/results/test_claim_stance_topic_pairs.json,sha256=Q4DP2sKyGgPqnerSfkLpmjGeYjUuEqdkaHLbs6GfADM,1910
|
|
1174
|
+
wisent/examples/scripts/results/test_cmmlu_evaluation.json,sha256=kD3YFj90X3YPDM8xYyHCVSZtINHXxa54n-W4jY6ad4I,3085
|
|
1175
|
+
wisent/examples/scripts/results/test_cmmlu_pairs.json,sha256=3CR0KaZccsrT3UBuox-T4xI3DGfZMWheFOT9rmSzOjU,1932
|
|
1176
|
+
wisent/examples/scripts/results/test_cnn_dailymail_evaluation.json,sha256=2xdFLCuLrI7Jhqu04QyhSPSjXPrwjKbPGhMV4kSjEl8,1710
|
|
1177
|
+
wisent/examples/scripts/results/test_cnn_dailymail_pairs.json,sha256=S1c_5HNR7YMIRy7z0fjMsRlCd5KhbPpmYXdgbl1WUIs,4727
|
|
1178
|
+
wisent/examples/scripts/results/test_cocoteros_es_evaluation.json,sha256=ahXlptFbAil4kf2iknnu8qPQG5Auklb_xbCeYIKjgSI,1181
|
|
1179
|
+
wisent/examples/scripts/results/test_cocoteros_es_pairs.json,sha256=x1XUZXKbohHuJcouIezJFi_1d_Ok8iGkuIVdtHVSo2A,385
|
|
1180
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_go_evaluation.json,sha256=T7k912Zo6CCp6KqP-uxv_wCltul3LdG3lzNdDNe8PwM,1144
|
|
1181
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_go_pairs.json,sha256=bE4f9bw6_bmzhDvBK7eUUcIp0YZdMUN4MLIABOjyReA,1271
|
|
1182
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_java_evaluation.json,sha256=eVluqR6mKYRYzBfUaY9vwg7vkwLkEtnBvHhcHxI9E2E,1146
|
|
1183
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_java_pairs.json,sha256=bE4f9bw6_bmzhDvBK7eUUcIp0YZdMUN4MLIABOjyReA,1271
|
|
1184
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_javascript_evaluation.json,sha256=t-DV2ml6tBNlxZCbZR85Mk6TLHPebVHp5XMF0mu4r0s,1152
|
|
1185
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_javascript_pairs.json,sha256=bE4f9bw6_bmzhDvBK7eUUcIp0YZdMUN4MLIABOjyReA,1271
|
|
1186
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_php_evaluation.json,sha256=wjaBRlJVk9FDqWLov8qhZSiGE5IX4ACCRfoycJqtpfs,1145
|
|
1187
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_php_pairs.json,sha256=bE4f9bw6_bmzhDvBK7eUUcIp0YZdMUN4MLIABOjyReA,1271
|
|
1188
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_python_evaluation.json,sha256=0vjJpfyI7YSiZk2cLDsRLKON62ekJ_yPpO-nLKBit6U,1148
|
|
1189
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_python_pairs.json,sha256=bE4f9bw6_bmzhDvBK7eUUcIp0YZdMUN4MLIABOjyReA,1271
|
|
1190
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_ruby_evaluation.json,sha256=iadmJqrfAQOrsoCGghFWi6uYfZWwQ2Y35Q4OQ-FK9m8,1146
|
|
1191
|
+
wisent/examples/scripts/results/test_codexglue_code_to_text_ruby_pairs.json,sha256=bE4f9bw6_bmzhDvBK7eUUcIp0YZdMUN4MLIABOjyReA,1271
|
|
1192
|
+
wisent/examples/scripts/results/test_coedit_gec_evaluation.json,sha256=oNucJ-YJW0OrWf8Jmlbf6ie3lCvM_HcYY1nddcvpIN4,1762
|
|
1193
|
+
wisent/examples/scripts/results/test_coedit_gec_pairs.json,sha256=80t6Cfu9be8TpPNaF0WOsRME6YbfLVyIBJyLGEuqM9o,687
|
|
1194
|
+
wisent/examples/scripts/results/test_cola_evaluation.json,sha256=4TOycnPImbfXliX05dKc8CUTW2MR8cKihmH0DRKR7OY,915
|
|
1195
|
+
wisent/examples/scripts/results/test_cola_pairs.json,sha256=8a5Dd84Qbeuy2a3CVpNfDwX4PGhYuw60j7YsM_5YwYE,249
|
|
1196
|
+
wisent/examples/scripts/results/test_commonsense_qa_evaluation.json,sha256=Dh2RPjf8vmF3Sqn4a83tPWUAX282KQ92NVj5UHmKpE0,889
|
|
1197
|
+
wisent/examples/scripts/results/test_commonsense_qa_pairs.json,sha256=GElMRlEYrHO2foSLcNMeYFZq_ILowHYV-ITl-RLkCcI,254
|
|
1198
|
+
wisent/examples/scripts/results/test_conala_evaluation.json,sha256=vFtN0a3pDm6tWG_0FxyYtKE4mX5Tl0NmvThkkP4YRfE,1153
|
|
1199
|
+
wisent/examples/scripts/results/test_conala_pairs.json,sha256=6HjJa_c-InK2aVa8TXyV5y0gwkldcKZdsw2PCBqrC88,307
|
|
1200
|
+
wisent/examples/scripts/results/test_concode_evaluation.json,sha256=jDU3G7zrHjeGL84AwJ9S6cgrFcQ6W0aHX_eJrrEN_gw,1126
|
|
1201
|
+
wisent/examples/scripts/results/test_concode_pairs.json,sha256=bE4f9bw6_bmzhDvBK7eUUcIp0YZdMUN4MLIABOjyReA,1271
|
|
1202
|
+
wisent/examples/scripts/results/test_copa_evaluation.json,sha256=7ds_h7aX0DclM-UT3kbFtfy5BTcC61-_GLj_PTa6tsk,1021
|
|
1203
|
+
wisent/examples/scripts/results/test_copa_pairs.json,sha256=9vZ5qZCZzMSLoLkSL2dW7JtvuaJ61INwowtMNJno6Mc,266
|
|
1204
|
+
wisent/examples/scripts/results/test_copal_id_evaluation.json,sha256=E00472fSXkOsrghGgu-ZOE3BLV5CK4L37_wHzDXDDps,1049
|
|
1205
|
+
wisent/examples/scripts/results/test_copal_id_pairs.json,sha256=oBZB1hq3BZkAudQyPIc2KAaaz7iQ6ZpWB1ceRkbjWgY,286
|
|
1206
|
+
wisent/examples/scripts/results/test_coqa_evaluation.json,sha256=Dbmjef69NmeTGM4F0b1FlCrSQES_scoQtxtUZXJpXqw,863
|
|
1207
|
+
wisent/examples/scripts/results/test_coqa_pairs.json,sha256=n2RU3_euUIwPzk1qCThIojrOo_z9JUIwNQwBQixs2mI,2348
|
|
1208
|
+
wisent/examples/scripts/results/test_coqcat_evaluation.json,sha256=i5pKT9mUk6iS2ubiq_mkW4utZIP15oVFqCK76a-adpE,978
|
|
1209
|
+
wisent/examples/scripts/results/test_coqcat_pairs.json,sha256=GWzqrA3XI7CK6VElef36IanFOOS7uJU-uKJSQ_8lQgg,1414
|
|
1210
|
+
wisent/examples/scripts/results/test_crows_pairs_evaluation.json,sha256=5LU5q4oHbHYTLzDz6_RL33WWzIimkY5FBI3jilhYNvo,2223
|
|
1211
|
+
wisent/examples/scripts/results/test_crows_pairs_pairs.json,sha256=DTMDM6kUAdiZb_osQFzBPJKW4ACLawbP91y_6O5pcE8,543
|
|
1212
|
+
wisent/examples/scripts/results/test_csatqa_evaluation.json,sha256=ZemydlxeR97I_NAsIWh4EvwfBInnJrqXYFP0PFD4VSU,2062
|
|
1213
|
+
wisent/examples/scripts/results/test_csatqa_pairs.json,sha256=EhaWZ6gF4iEZZiPdgmA57FT3Rjxm0BX69wUOHlv7NPY,16738
|
|
1214
|
+
wisent/examples/scripts/results/test_cycle_letters_evaluation.json,sha256=ynLm4vDWMSJKwbl-Xsemgz1m5Z1aEdQ3MB4n28RJONI,804
|
|
1215
|
+
wisent/examples/scripts/results/test_cycle_letters_pairs.json,sha256=Rn_PkojOpuiIDTHFBirHdfeia101K0EBo_GSpPPM1Ew,125
|
|
1216
|
+
wisent/examples/scripts/results/test_darija_bench_evaluation.json,sha256=mE-UlpD4_bpnlQt-rYW-4S4hS_HXOKpjup-f4pj20hw,2599
|
|
1217
|
+
wisent/examples/scripts/results/test_darija_bench_pairs.json,sha256=td8oZ9RFS5qeVx3gNdf6bgo3yjWGd6Tkz7jOa-8zByg,1711
|
|
1218
|
+
wisent/examples/scripts/results/test_darijahellaswag_evaluation.json,sha256=FGtUWgfmQsswahqXkE4LCSsHZz07gdJOiNdhkN9yCf4,2249
|
|
1219
|
+
wisent/examples/scripts/results/test_darijahellaswag_pairs.json,sha256=-buMWLczhUuoI2nWlszur0lhC7QkIh3jKfDMTKXpiYY,1075
|
|
1220
|
+
wisent/examples/scripts/results/test_darijammlu_evaluation.json,sha256=pPFB5GyxpL5la7L4g1fliScubibw7njgK1-oLvouaL8,4884
|
|
1221
|
+
wisent/examples/scripts/results/test_darijammlu_pairs.json,sha256=NGGpYIdGsT01HfjC0zYGOwBZWb8TYVEOk0s2WzOlCCU,4938
|
|
1222
|
+
wisent/examples/scripts/results/test_dbpedia_14_evaluation.json,sha256=9IKJ7skQ9IChdGin5YH2C8FD4dRZwb6zYG4IypiZ4eA,889
|
|
1223
|
+
wisent/examples/scripts/results/test_dbpedia_14_pairs.json,sha256=5DCuHnpY4v9-JNSoQwOh-UG9bFBNF2CJi9RBqVTIyBc,443
|
|
1224
|
+
wisent/examples/scripts/results/test_drop_evaluation.json,sha256=O0SvU8t-t35INLizNmLgYZoc3ToeySPEvRnn6VmRgqM,927
|
|
1225
|
+
wisent/examples/scripts/results/test_drop_pairs.json,sha256=_68MgtdAUN5y1_zQ5IdFGPXRjcbJhJcikbWU4FuQO9s,1140
|
|
1226
|
+
wisent/examples/scripts/results/test_ds1000_evaluation.json,sha256=5NTDhZcD_ni6XuXrxn3qI9l7JJ4FwSHiCujMpvURYTM,1314
|
|
1227
|
+
wisent/examples/scripts/results/test_ds1000_pairs.json,sha256=56SnkofYQiXDXYGjWYbKxOfUhW0hSew4GyZuaGgE9e8,1449
|
|
1228
|
+
wisent/examples/scripts/results/test_egyhellaswag_evaluation.json,sha256=Mnndav8dU-XSQSuhi62jix-O2XCTj50nJ6XW41Yrt_o,1776
|
|
1229
|
+
wisent/examples/scripts/results/test_egyhellaswag_pairs.json,sha256=p-3YQa3I76CEMxbrky9C5i1JqGLV_fB7-bMU-qAuOE0,605
|
|
1230
|
+
wisent/examples/scripts/results/test_egymmlu_evaluation.json,sha256=WoxCCYTlcCfJlSWlrYWnNNi_yjnJSBvJO3GWQCdYnWI,2817
|
|
1231
|
+
wisent/examples/scripts/results/test_egymmlu_pairs.json,sha256=aA_nRrn1UDmGT7MkXvgijhP9sa9Rnyfpfs-V2GJhHSE,1914
|
|
1232
|
+
wisent/examples/scripts/results/test_epec_koref_bin_evaluation.json,sha256=uptYxyqdNCvVAMXKVBzOe7MD-iagHcaPDDHzBPkeIJU,875
|
|
1233
|
+
wisent/examples/scripts/results/test_epec_koref_bin_pairs.json,sha256=_D4-SAq-KYq4Dh7Sz4Je_grkJs2A16XhYRcjWHMDBfM,395
|
|
1234
|
+
wisent/examples/scripts/results/test_eq_bench_evaluation.json,sha256=B-AqxREzlu0ECD_YtS---LdSKolwu70wwcu7AKE8HeY,1203
|
|
1235
|
+
wisent/examples/scripts/results/test_eq_bench_pairs.json,sha256=phvyjATGYJ1epmiCCdOBPxN1l_AO792udzCuAiZnWqY,1364
|
|
1236
|
+
wisent/examples/scripts/results/test_escola_evaluation.json,sha256=JKPRlQVkuaDLsLKzI5XxN4gwyOuJH04Qh37QDMB0TmA,939
|
|
1237
|
+
wisent/examples/scripts/results/test_escola_pairs.json,sha256=SiA_WBXRzwxUnQ2_j56d3t3MJyt7iklZguDmWl06rlI,177
|
|
1238
|
+
wisent/examples/scripts/results/test_ethics_cm_evaluation.json,sha256=MmzSjNpXgo2SRv37SDy9jm6uCP0C2gF73Z9B-oTVsdk,836
|
|
1239
|
+
wisent/examples/scripts/results/test_ethics_cm_pairs.json,sha256=00wwW87OJi4_RUt3AizJR6lMLTPYOkKhGzy7uwYlmhk,177
|
|
1240
|
+
wisent/examples/scripts/results/test_ethos_binary_evaluation.json,sha256=XeHT77YVOvGVUtIqu73B6FuxkxsdtWh8-C_U-5dYhsQ,939
|
|
1241
|
+
wisent/examples/scripts/results/test_ethos_binary_pairs.json,sha256=mfqI7rEbpZlnyMCufGG5QcCyq4iOSEJHZiKEbY7Mm7M,449
|
|
1242
|
+
wisent/examples/scripts/results/test_eus_exams_es_evaluation.json,sha256=eCeeKM0EKjAuyYuQo8iW7wDL0XlVuTFPl3n6cE5tVZw,4112
|
|
1243
|
+
wisent/examples/scripts/results/test_eus_exams_es_pairs.json,sha256=Bn3Xm1qJseSwhdAT8j3F4vu3u3K6e3smdMimdorX_JM,2025
|
|
1244
|
+
wisent/examples/scripts/results/test_eus_exams_evaluation.json,sha256=aMWHOmI1qOEPnUadmadu28pYLb4u-_Kxj2OHfHgXs4w,4109
|
|
1245
|
+
wisent/examples/scripts/results/test_eus_exams_pairs.json,sha256=Bn3Xm1qJseSwhdAT8j3F4vu3u3K6e3smdMimdorX_JM,2025
|
|
1246
|
+
wisent/examples/scripts/results/test_eus_proficiency_evaluation.json,sha256=cfm_4e8ZCArtf4t6c55JOSwgJ_N7EELQ9PsZ9zhq3T4,922
|
|
1247
|
+
wisent/examples/scripts/results/test_eus_proficiency_pairs.json,sha256=6JzoYrt_c26GokB3znOT7TgvY0O1EWLT_kqmKy26oJg,211
|
|
1248
|
+
wisent/examples/scripts/results/test_eus_reading_evaluation.json,sha256=IH05FVYgcDTVuTqddbdgzvmvBRIr1s4AuFjzqIz5x30,1385
|
|
1249
|
+
wisent/examples/scripts/results/test_eus_reading_pairs.json,sha256=PKa11xuBbOcI6FpAREUMesxjfIDhlSZmRdtbogECZLo,498
|
|
1250
|
+
wisent/examples/scripts/results/test_eus_trivia_evaluation.json,sha256=urYqWTWSYJcoaS1o1QRW_6e7vWfYRKsXUivmMxZSR4E,913
|
|
1251
|
+
wisent/examples/scripts/results/test_eus_trivia_pairs.json,sha256=sB3Pn-0utMG9Z1GrfATsLszoWobvY_Po51r3g2sRvxU,214
|
|
1252
|
+
wisent/examples/scripts/results/test_evalita-mp_evaluation.json,sha256=5MS0hwskRstAR2BhPOg-4-yl5J7AqFG8WfABkb4mYXE,1904
|
|
1253
|
+
wisent/examples/scripts/results/test_evalita-mp_pairs.json,sha256=HjC70ItXgdB5iE0LyEILkNb7JHsOrFsb3AGNhZc44zs,1785
|
|
1254
|
+
wisent/examples/scripts/results/test_evalita-sp_sum_task_fp-small_p1_evaluation.json,sha256=QfzE-7Sltc5ZiAn0HH66dgLNwLFtR3SgV_PW0p_FHiM,2869
|
|
1255
|
+
wisent/examples/scripts/results/test_evalita-sp_sum_task_fp-small_p1_pairs.json,sha256=BUTzfqVMdF-_emOOOn4cpvIEhiO4ggjs9tnak--DTNw,3268
|
|
1256
|
+
wisent/examples/scripts/results/test_evalita_LLM_evaluation.json,sha256=XAv0b6ZsY3FQTfePmAwZctCZKQjaTedZ5aDDMeGr_gU,1905
|
|
1257
|
+
wisent/examples/scripts/results/test_evalita_LLM_pairs.json,sha256=HjC70ItXgdB5iE0LyEILkNb7JHsOrFsb3AGNhZc44zs,1785
|
|
1258
|
+
wisent/examples/scripts/results/test_fda_evaluation.json,sha256=aP74kYT6BOD2qM05L4kRnAYuVw-tZvitE9caeG8e__M,998
|
|
1259
|
+
wisent/examples/scripts/results/test_fda_pairs.json,sha256=XZTBelsFlYkB0-valzK6W96-p-TE-8nlFq7zVhMCSa8,8182
|
|
1260
|
+
wisent/examples/scripts/results/test_financial_tweets_evaluation.json,sha256=SDmml0B30P1RSZiQY502QTV7fXyUH_YHlE0YJjy7Xm0,921
|
|
1261
|
+
wisent/examples/scripts/results/test_financial_tweets_pairs.json,sha256=JcsMYuLOrJ550nh_zrN8v2MxYNvIhtvReHFwD0OPVek,669
|
|
1262
|
+
wisent/examples/scripts/results/test_fld_evaluation.json,sha256=aiEXO4lqEdjtTzY4hFbFTH2CPZXeDPfKfme87WacMhE,842
|
|
1263
|
+
wisent/examples/scripts/results/test_fld_pairs.json,sha256=1L5jg_v5WHopxNJYpEplZXRowQ4hH3bKT3G2TMnW-CQ,1629
|
|
1264
|
+
wisent/examples/scripts/results/test_flores_evaluation.json,sha256=vDkapQOxi5g2qal8IEGF08szTWxM8ATQHRKiIfi4b-s,3915
|
|
1265
|
+
wisent/examples/scripts/results/test_flores_pairs.json,sha256=HRmZsjTuME55GRtXP5_KBPsxojNefPQPRs4l_QWc2_M,1910
|
|
1266
|
+
wisent/examples/scripts/results/test_freebase_evaluation.json,sha256=N3kgEsfL-mkSn4H6nyX5wUVlE8GSb7-H-X8XKnhBv3o,1048
|
|
1267
|
+
wisent/examples/scripts/results/test_freebase_pairs.json,sha256=29RyCFZUDEVuo9okvNWGWD4wvF-pvnTtV33F2gkV8wI,293
|
|
1268
|
+
wisent/examples/scripts/results/test_french_bench_evaluation.json,sha256=fleGxHmOfhe1_pVl_snl8CMfrqj4_QiL4UpIjQBPIHQ,2677
|
|
1269
|
+
wisent/examples/scripts/results/test_french_bench_pairs.json,sha256=Vb-FUJaL-wTV0cEgqpVmfJwegmeZniRRqirwT3HUuNk,1258
|
|
1270
|
+
wisent/examples/scripts/results/test_galcola_evaluation.json,sha256=3lMUcse5omV9hk17btflMTZnOq7y1TtAsr_krXbuQvw,982
|
|
1271
|
+
wisent/examples/scripts/results/test_galcola_pairs.json,sha256=D4ECFn7QtOstIXignhIHP9acClIrh0vpe_oKZdxGbNE,243
|
|
1272
|
+
wisent/examples/scripts/results/test_galician_bench_evaluation.json,sha256=3KKwdau0a-5knxgX6kkVrlPlxueTRwRBNravHkmwr2E,5047
|
|
1273
|
+
wisent/examples/scripts/results/test_galician_bench_pairs.json,sha256=srtHtUgyzmAOEqnJhWJisaKhSTRJYw4U7U6jNIQgmQE,2886
|
|
1274
|
+
wisent/examples/scripts/results/test_glianorex_evaluation.json,sha256=6CLnzisor1mMqlxpLvTgYTXQmY3UhHsk7Nf0zxb7UO8,1006
|
|
1275
|
+
wisent/examples/scripts/results/test_glianorex_pairs.json,sha256=mqwrGlAwyWtMzftOfO_4HQbKVGNSppP_iDbMU5AvGZQ,658
|
|
1276
|
+
wisent/examples/scripts/results/test_global_mmlu_evaluation.json,sha256=EQiG1RkHTLvR4VBEXiv8b5vHXOq_PajfNIij7_YmT7c,2133
|
|
1277
|
+
wisent/examples/scripts/results/test_global_mmlu_pairs.json,sha256=spLwFJkOGruG-i1Csn6RCpKLr0-a8CVpQg4mzvO0JFs,826
|
|
1278
|
+
wisent/examples/scripts/results/test_glue_evaluation.json,sha256=xXKAFKDXmg0xaCj369bPD570sNIjOeLOVtTbSUkABvA,1600
|
|
1279
|
+
wisent/examples/scripts/results/test_glue_pairs.json,sha256=mymkijh1I9vEaoiIzq39Lm0L73UShIxQFBMcayoB7j4,628
|
|
1280
|
+
wisent/examples/scripts/results/test_gpqa_evaluation.json,sha256=2fEW-8wj4TDoHXhCas_YPZ4Z_N7xJiAzjXry7C3-lCk,1840
|
|
1281
|
+
wisent/examples/scripts/results/test_gpqa_pairs.json,sha256=y-TOv-Lo40zSpXVo4MXPGZTkIu_u3WNs4FZjjsK6upc,1084
|
|
1282
|
+
wisent/examples/scripts/results/test_gpt3_translation_benchmarks_evaluation.json,sha256=IdcK2tguDuY_WUWj4odEan4Oqjy_FVUfiPJiiP7JLxc,4101
|
|
1283
|
+
wisent/examples/scripts/results/test_gpt3_translation_benchmarks_pairs.json,sha256=3dOdaSTHcX0ccOVYYIWkpq687tJwFL1eIFc6EHoIKww,1419
|
|
1284
|
+
wisent/examples/scripts/results/test_groundcocoa_evaluation.json,sha256=rV_veFpVK3OImuZyzulwmph6NAdy7pI156SfE_dOKeM,988
|
|
1285
|
+
wisent/examples/scripts/results/test_groundcocoa_pairs.json,sha256=CirRCcb2q2pmIu8e6ANHBuubuvtj2shGc0zDjg0lBZQ,5350
|
|
1286
|
+
wisent/examples/scripts/results/test_gsm8k_evaluation.json,sha256=_nNDLwwfjqOFihfMuv0O3ZF_Ew2_0k_D-2o-d_DcZ5o,867
|
|
1287
|
+
wisent/examples/scripts/results/test_gsm8k_pairs.json,sha256=yxW6_k2XylyEqPMPhZhwxyCnw-vwCYBedSkx1o5Zeg4,403
|
|
1288
|
+
wisent/examples/scripts/results/test_haerae_evaluation.json,sha256=uK5ol-97fU29DkV0Vrb9L_4KScnSJzGOP8UU2Q8MFrg,2215
|
|
1289
|
+
wisent/examples/scripts/results/test_haerae_pairs.json,sha256=ltoQ282NCuvZlRnxBxrSc0sp1XqbRDGDJVv1PlMIMHQ,1468
|
|
1290
|
+
wisent/examples/scripts/results/test_headqa_evaluation.json,sha256=0k3HEcKU7qSyNN7W79tHDxGZopq_kqzbRMFfUtmx1K4,1029
|
|
1291
|
+
wisent/examples/scripts/results/test_headqa_pairs.json,sha256=rGY9CijTDzGGL85pxKZXZe_M-Px_E-uGOPOxqU26_M8,351
|
|
1292
|
+
wisent/examples/scripts/results/test_hellaswag_evaluation.json,sha256=_-iilhXZXRsaWp7LlJMFTc-Qmh2xCK4GOEB5a2mB3f0,1074
|
|
1293
|
+
wisent/examples/scripts/results/test_hellaswag_pairs.json,sha256=rbSvCblmQGvUgw-kZL03kidiCfiTIVfLnt_0yvmSoZI,312
|
|
1294
|
+
wisent/examples/scripts/results/test_hendrycks_ethics_evaluation.json,sha256=0wHr0vhBm2i8_3lgqhGrwcM65YuPgs33G6cxVDF609w,1598
|
|
1295
|
+
wisent/examples/scripts/results/test_hendrycks_ethics_pairs.json,sha256=uBKdLhk00i6DAAtQut1B0tqxx19kiPBEjlGBxFTfTwM,458
|
|
1296
|
+
wisent/examples/scripts/results/test_hendrycks_math_evaluation.json,sha256=s2orwBnwfMYspX3_GD3FO5e8J2ZuEwdlQDOrMKCbgMU,1665
|
|
1297
|
+
wisent/examples/scripts/results/test_hendrycks_math_pairs.json,sha256=MyZHhACbkn0PkMmAChQh3wIXOWrXpWYiNyLpmtSgo_4,445
|
|
1298
|
+
wisent/examples/scripts/results/test_histoires_morales_evaluation.json,sha256=5e8X1La9oLlLkkRRnU5zoSuf8CuGondBEwXM7pWIJSo,1396
|
|
1299
|
+
wisent/examples/scripts/results/test_histoires_morales_pairs.json,sha256=g1rM_DrRSeTwOG7SDGknO2W6Nyqb5NSS0vHFQzxqpjA,565
|
|
1300
|
+
wisent/examples/scripts/results/test_hmmt_evaluation.json,sha256=8XB_yyMUQOuzkYX__vmQSos4vDNX6Yt_MaDc9anyX1k,867
|
|
1301
|
+
wisent/examples/scripts/results/test_hmmt_feb_2025_evaluation.json,sha256=OIDOuto4Qm9HCBCjX3CVUhDIBQp6Ml4SeJ4yAzlraWI,876
|
|
1302
|
+
wisent/examples/scripts/results/test_hmmt_feb_2025_pairs.json,sha256=bfe5YJsR1WMdMNkaMjYBGxU8hqZaKLguLY9afvlCfjA,234
|
|
1303
|
+
wisent/examples/scripts/results/test_hmmt_pairs.json,sha256=bfe5YJsR1WMdMNkaMjYBGxU8hqZaKLguLY9afvlCfjA,234
|
|
1304
|
+
wisent/examples/scripts/results/test_hrm8k_evaluation.json,sha256=6HYYFhWhpzq2Y0DYk6NfJOmvHFl_shwL_x7SCbx46CY,1875
|
|
1305
|
+
wisent/examples/scripts/results/test_hrm8k_pairs.json,sha256=lZN6MfCBQ6Bq1GKVsaNE-CcLbT7TmwkzVcw_LXMFlQU,839
|
|
1306
|
+
wisent/examples/scripts/results/test_humaneval_evaluation.json,sha256=HC7TbsZN7SrjpHUBpdEWM35wBQj6jGLlXJnovjcjyl4,4167
|
|
1307
|
+
wisent/examples/scripts/results/test_humaneval_pairs.json,sha256=qZXOhERJoMGYk8J5EFoKLyKYu49_l5OealvoML-RWmo,1541
|
|
1308
|
+
wisent/examples/scripts/results/test_humaneval_plus_evaluation.json,sha256=ldl3iyWRY4Sc7WEFj-hJ7BQD0-QRBbgHq9l4JGPF8GA,4172
|
|
1309
|
+
wisent/examples/scripts/results/test_humaneval_plus_pairs.json,sha256=qZXOhERJoMGYk8J5EFoKLyKYu49_l5OealvoML-RWmo,1541
|
|
1310
|
+
wisent/examples/scripts/results/test_ifeval_evaluation.json,sha256=gOtlvCuVrfBgSJ73Xa5Yfjuv3z5U7aI17SHk13fg4F0,1104
|
|
1311
|
+
wisent/examples/scripts/results/test_ifeval_pairs.json,sha256=TCi8sdQ9kbjF1Ob9MQaJf10JBTpUlYmN_BrudTWclHk,534
|
|
1312
|
+
wisent/examples/scripts/results/test_instruct_humaneval_evaluation.json,sha256=UMpGUa1qXuXRocy6eBl2swB3D_R9fwwrVchzb-uFCHY,2214
|
|
1313
|
+
wisent/examples/scripts/results/test_instruct_humaneval_pairs.json,sha256=-XPga07HK8oiOMD1i22smzUp1KAI5xqg81YQf6ofdi4,1836
|
|
1314
|
+
wisent/examples/scripts/results/test_inverse_scaling_evaluation.json,sha256=oHUzSzzhF-Lfb91VZX-r869qacnyifJxGcEswDrjOfg,1587
|
|
1315
|
+
wisent/examples/scripts/results/test_inverse_scaling_hindsight_neglect_10shot_evaluation.json,sha256=NJN-QphkJMpBo5FD8c6vMHWVwG7MJjq621fPSERLZaA,897
|
|
1316
|
+
wisent/examples/scripts/results/test_inverse_scaling_hindsight_neglect_10shot_pairs.json,sha256=dmiHcLpmP90-KaWrHXWcRFaLI51cl7wKJBS1OVvkBXk,3043
|
|
1317
|
+
wisent/examples/scripts/results/test_inverse_scaling_pairs.json,sha256=BGaTxMETM5oGL24Mukz2_EgsUmGwHphm5tJ5zrnh7oo,1973
|
|
1318
|
+
wisent/examples/scripts/results/test_iwslt2017-ar-en_evaluation.json,sha256=JUjNmv-0VSb3NeX2v5oYNCijgvSDiYsFawf7G-h0Bx4,2307
|
|
1319
|
+
wisent/examples/scripts/results/test_iwslt2017-ar-en_pairs.json,sha256=mHj6LEr0usvxdt5K4rbi64c0ZqTMmFi4i57JjDpCesw,1414
|
|
1320
|
+
wisent/examples/scripts/results/test_iwslt2017-en-ar_evaluation.json,sha256=VMZGv37Ji7PjjmlliajLVehPWbafzkq0TGzsjXQi_zA,2307
|
|
1321
|
+
wisent/examples/scripts/results/test_iwslt2017-en-ar_pairs.json,sha256=hnCYmmkPDL9fVosNOMKtwKuz64Zik26kn5z5NT__GVE,1414
|
|
1322
|
+
wisent/examples/scripts/results/test_japanese_leaderboard_evaluation.json,sha256=yHx8LLslijfskvgapcNb8DX5tTmz_rKNUo3uYxliEL8,1901
|
|
1323
|
+
wisent/examples/scripts/results/test_japanese_leaderboard_pairs.json,sha256=sQdfqPcDetEMeHbOAeyt5r5WxVC1h3b2Iat4ob7rLNo,613
|
|
1324
|
+
wisent/examples/scripts/results/test_jsonschema_bench_evaluation.json,sha256=Vnp1Ha2PyH-VA3IasJtKQ4w9AJKSZldqcV9voMAP980,893
|
|
1325
|
+
wisent/examples/scripts/results/test_jsonschema_bench_pairs.json,sha256=rTNcw8zRB8sino6WuO6P2lhKEiRG2VHXSh-a2mhJMdc,19480
|
|
1326
|
+
wisent/examples/scripts/results/test_kbl_evaluation.json,sha256=8kGpFhxjO8x_HRSvAg6-WbTez2E8A51ES0i1b0MOCCM,1980
|
|
1327
|
+
wisent/examples/scripts/results/test_kbl_pairs.json,sha256=LV-Ua4yLlHsewRWlSYsZwTQsd9Gph9iJlDyrgPTTGdo,1803
|
|
1328
|
+
wisent/examples/scripts/results/test_kmmlu_evaluation.json,sha256=_ipQsnbnlMqNAcya6x_1yaEukdv75GShBsduslKP6js,4367
|
|
1329
|
+
wisent/examples/scripts/results/test_kmmlu_pairs.json,sha256=k4ZJBgxx-gYI0LFrCHfO_b-8eCA2Q6l1jX-lua53Nxk,2216
|
|
1330
|
+
wisent/examples/scripts/results/test_kobest_evaluation.json,sha256=GazAN4Lllr17o1Xyy4KXpGSpLjxAqO_zGFsLTd-l88k,2251
|
|
1331
|
+
wisent/examples/scripts/results/test_kobest_pairs.json,sha256=IWkCy1MdcVYoEk6vKaRGG5rERbsUgig5B-H_w4Bojsg,1855
|
|
1332
|
+
wisent/examples/scripts/results/test_kormedmcqa_evaluation.json,sha256=j85PbuvVbOUZCiROCx5s6xGeaDEqGvJmY-zFzC-To-Y,1221
|
|
1333
|
+
wisent/examples/scripts/results/test_kormedmcqa_pairs.json,sha256=1zItWyoLhYMnSx5dEUSm62Wl47PTgM7IFA0vabZ7yNU,1133
|
|
1334
|
+
wisent/examples/scripts/results/test_lambada_cloze_evaluation.json,sha256=gvT47FFYyM9Z9Nh_Uko-hAaOd63B82MjHLkdT_hvPas,887
|
|
1335
|
+
wisent/examples/scripts/results/test_lambada_cloze_pairs.json,sha256=8RCMdUcaJRjHeayPWYdV_3uXiurW6zzwUWtAvAIzsyU,458
|
|
1336
|
+
wisent/examples/scripts/results/test_lambada_evaluation.json,sha256=beUxcF4FDT7j8ZqXMkiUOJjfR7FvN1oMwmJjO7SLK7o,881
|
|
1337
|
+
wisent/examples/scripts/results/test_lambada_multilingual_evaluation.json,sha256=DNICi_fRYNDqTwCKkTViMh5u2Tb9r2fMCvYAFDAKUhw,1648
|
|
1338
|
+
wisent/examples/scripts/results/test_lambada_multilingual_pairs.json,sha256=J_wDReRhKlfTCNowV9xqgU0ko4uewXp6VVEJhVvt5ro,930
|
|
1339
|
+
wisent/examples/scripts/results/test_lambada_multilingual_stablelm_evaluation.json,sha256=qDYpgEeyaeoQhDYWC9Vnb8n6-sQ6wl9Ii4ojl2maaJ0,1674
|
|
1340
|
+
wisent/examples/scripts/results/test_lambada_multilingual_stablelm_pairs.json,sha256=4CTyzbFCDFjq6wIzR30iL_TBjx9yYpm5rj1ZjTj7thw,827
|
|
1341
|
+
wisent/examples/scripts/results/test_lambada_openai_evaluation.json,sha256=tObl_eY7drZaVDK6GDdRaqN7clgcD45iBX5YozJK-UY,886
|
|
1342
|
+
wisent/examples/scripts/results/test_lambada_openai_pairs.json,sha256=xMGsuzVg68MYRlY390HChWBK0BcxeWiwAK4hobQ0NPc,451
|
|
1343
|
+
wisent/examples/scripts/results/test_lambada_pairs.json,sha256=wN49Lxh6CB442qrLeP2Jo2-F8clJtx1C5W9GJzXIt2o,452
|
|
1344
|
+
wisent/examples/scripts/results/test_lambada_standard_evaluation.json,sha256=jSewuBV9w-EIab_RpOIwFn6RxuLb4XKCJYa4ljwjfHk,885
|
|
1345
|
+
wisent/examples/scripts/results/test_lambada_standard_pairs.json,sha256=zh9Ru3R849fBwppp8UZoj7BscDjyxW8qfeyOtAroyzs,589
|
|
1346
|
+
wisent/examples/scripts/results/test_leaderboard_evaluation.json,sha256=Tpck6aCefnexgobcfXM3ji8hbwMJ0yxAfeVBZBEzjFA,1917
|
|
1347
|
+
wisent/examples/scripts/results/test_leaderboard_pairs.json,sha256=8B0GFqsopf9uyW0rf_M2-Tv7ufxm1NB5FzRt5AqOTT0,872
|
|
1348
|
+
wisent/examples/scripts/results/test_libra_evaluation.json,sha256=XvgFoua3CpIO7PNBVzKFgji9rsU83YIE4CHNUJW2jZg,2485
|
|
1349
|
+
wisent/examples/scripts/results/test_libra_pairs.json,sha256=QA0Vxch6FIOKpd17z1ytL9TTA2vGnp-qYFtmk3ePu7M,880379
|
|
1350
|
+
wisent/examples/scripts/results/test_lingoly_evaluation.json,sha256=i3to6V4qYutI9nyjx_7lowdPeTE6VDxbn0X4MRCE-KE,1114
|
|
1351
|
+
wisent/examples/scripts/results/test_lingoly_pairs.json,sha256=2iAkMsykDno2K4ZkCSn2QdSbBqxXtk6S68jtz2gtjfU,3397
|
|
1352
|
+
wisent/examples/scripts/results/test_livecodebench_evaluation.json,sha256=vwTMrFUwRRSp-P4eisuiyUzHfwUWSC8xpBk-d2YX7h0,3481
|
|
1353
|
+
wisent/examples/scripts/results/test_livecodebench_pairs.json,sha256=W3P0-akjy_5kbetws0yAvnGUSK-7BNk99w5affkdQo4,2463
|
|
1354
|
+
wisent/examples/scripts/results/test_livemathbench_cnmo_en_evaluation.json,sha256=xigdJBdMptJZ-6aLe8eDmEH6T7w7iTjYEqX09Xtaz-s,900
|
|
1355
|
+
wisent/examples/scripts/results/test_livemathbench_cnmo_en_pairs.json,sha256=McXMGtCz2UD2QUXqylo1sTf-a6iAmoo5qLIthtNHVWo,260
|
|
1356
|
+
wisent/examples/scripts/results/test_livemathbench_cnmo_zh_evaluation.json,sha256=BkFqTOBtPM9fuSsEPV-4mNzWhV2vWJmBu5YRGGqCeRU,956
|
|
1357
|
+
wisent/examples/scripts/results/test_livemathbench_cnmo_zh_pairs.json,sha256=rvpcmohMC5eBnL5im_-JdUPRHh7XFFJZ1gt1JPM0osk,272
|
|
1358
|
+
wisent/examples/scripts/results/test_llama_evaluation.json,sha256=iDVGx1269b7B8W2fdYPX0LPSnS_V-WOZZR8jnuFv9ns,1062
|
|
1359
|
+
wisent/examples/scripts/results/test_llama_pairs.json,sha256=i_ZNQATvJAvvUh3IyoEZaVz2PPK3e9qakj96jTkX4jw,415
|
|
1360
|
+
wisent/examples/scripts/results/test_logiqa2_evaluation.json,sha256=Za93i3cFpej8eA1jn62rHi1LMKQPe88QSm6_vC_Hy5o,1402
|
|
1361
|
+
wisent/examples/scripts/results/test_logiqa2_pairs.json,sha256=KVIfTmhyZdttg2ryAgrl-oO16VK4_kvSZM2WeL2bwig,908
|
|
1362
|
+
wisent/examples/scripts/results/test_logiqa_evaluation.json,sha256=C8iRLhxAXwTFy8ptdIZx-Lwie13NZ4-Sewa51BRDSfo,1691
|
|
1363
|
+
wisent/examples/scripts/results/test_logiqa_pairs.json,sha256=elVOWcg1jhq0dIF7RLnDA2MDK4NN_FnPZN-grM343qA,1272
|
|
1364
|
+
wisent/examples/scripts/results/test_m_mmlu_evaluation.json,sha256=PclJkewbHm3_Rw5qTC8NfOT2NgERpZmuYbn2oyzJV_g,3301
|
|
1365
|
+
wisent/examples/scripts/results/test_m_mmlu_pairs.json,sha256=wAAzpckbG1ZtJzULClaMmzjMb58vCexgJy04I3en6po,6160
|
|
1366
|
+
wisent/examples/scripts/results/test_mastermind_evaluation.json,sha256=f69woRnb2_yIiKm6JNTpM17z3YkisTqiYgwTgvv8Fzc,1846
|
|
1367
|
+
wisent/examples/scripts/results/test_mastermind_pairs.json,sha256=nZiJ_cPmiPfl799XB7t5vejrct1fZYCG_dWvHiGbKS0,1778
|
|
1368
|
+
wisent/examples/scripts/results/test_math500_evaluation.json,sha256=F8FHRWhePZGYu3k29BmDJEmwt14LDIfEkWk2yPOv6p8,1070
|
|
1369
|
+
wisent/examples/scripts/results/test_math500_pairs.json,sha256=h-g2WQ-86hoGHOly0_DD2fG7FJre0P8NUihhOI5NWkE,376
|
|
1370
|
+
wisent/examples/scripts/results/test_math_evaluation.json,sha256=-kzF9nksoli-AaMJRyvSw5qJK8c65XDcWZbQp3gABso,1067
|
|
1371
|
+
wisent/examples/scripts/results/test_math_pairs.json,sha256=h-g2WQ-86hoGHOly0_DD2fG7FJre0P8NUihhOI5NWkE,376
|
|
1372
|
+
wisent/examples/scripts/results/test_mathqa_evaluation.json,sha256=X-1i9Lz5YUamprD48dYZE5K9w9YjzyMutU0EpciTVWQ,867
|
|
1373
|
+
wisent/examples/scripts/results/test_mathqa_pairs.json,sha256=faRg8WeEewq6vGd6jl2p7vcrJ_tDUqAKjuVLyZ5bPYg,296
|
|
1374
|
+
wisent/examples/scripts/results/test_mbpp_evaluation.json,sha256=u52WM6boj2wKelvH8mdmDE5OTwvA3at1EdfQDGct6vk,2619
|
|
1375
|
+
wisent/examples/scripts/results/test_mbpp_pairs.json,sha256=Gkc2VORp4d11JV7OGQBSumeMSr6svf69Hc74ZiluKvY,840
|
|
1376
|
+
wisent/examples/scripts/results/test_mbpp_plus_evaluation.json,sha256=s1lIEhaqSsSnzCk-pXeuq4lK-PM-FvO9g4l67KI4SwA,2624
|
|
1377
|
+
wisent/examples/scripts/results/test_mbpp_plus_pairs.json,sha256=Gkc2VORp4d11JV7OGQBSumeMSr6svf69Hc74ZiluKvY,840
|
|
1378
|
+
wisent/examples/scripts/results/test_mc_taco_evaluation.json,sha256=1Y5oIUdsosSJoYLmur7l11epNgIE9EiGXGdTtx3wxOs,866
|
|
1379
|
+
wisent/examples/scripts/results/test_mc_taco_pairs.json,sha256=WPbTNx9ADn2LSU-P1jAs7-iarNG0Z-uCdHqwF3tnDzU,343
|
|
1380
|
+
wisent/examples/scripts/results/test_med_concepts_qa_evaluation.json,sha256=ouFMxeZaFneONKOFiCabnHUhi6GK5nqN0v0BLYwj_W4,1577
|
|
1381
|
+
wisent/examples/scripts/results/test_med_concepts_qa_pairs.json,sha256=QRn1FPe4dRcoE1E1DzIlg7hCqnVdeBZiXSeteJW44Kk,926
|
|
1382
|
+
wisent/examples/scripts/results/test_meddialog_evaluation.json,sha256=f4avjS3zlNiMbev2Bji9kVycJkXbOMEfC5LyTMvlDfs,1224
|
|
1383
|
+
wisent/examples/scripts/results/test_meddialog_pairs.json,sha256=Az1EqGbgyi-w-rLU0jdBN1KF1njqonbSowiY0dahbTU,2211
|
|
1384
|
+
wisent/examples/scripts/results/test_mediqa_qa2019_evaluation.json,sha256=Fvfph4eUhNd2jfrUjwFC71tl_b6SuG_yWvztP33maYQ,24657
|
|
1385
|
+
wisent/examples/scripts/results/test_mediqa_qa2019_pairs.json,sha256=7y_9xrpYvGk1D8A3oz6TLB481lH0vaemYR0WTDSCYvg,6176
|
|
1386
|
+
wisent/examples/scripts/results/test_medmcqa_evaluation.json,sha256=P__zH8WsY20aBMZvENbSzvv4dsm6kCsnjbMsx0BkpO0,1235
|
|
1387
|
+
wisent/examples/scripts/results/test_medmcqa_pairs.json,sha256=-RPjq5ws9WEHvDuA2kKb7PcLt_xO_pKHOmX9ZFo0vOA,429
|
|
1388
|
+
wisent/examples/scripts/results/test_medqa_evaluation.json,sha256=7Vit2oW56PIRdB7KIuc-0De1ViUZy_0mnCRxbK6krhw,918
|
|
1389
|
+
wisent/examples/scripts/results/test_medqa_pairs.json,sha256=r7IpFV92Er1y8qQG1uP3AgqMx43lKlHmS7ghSHuWuk4,580
|
|
1390
|
+
wisent/examples/scripts/results/test_medtext_evaluation.json,sha256=os6rwCha3MxiLHtb2K4_d2Pw8ETjsUUIz9gluDagaC8,3320
|
|
1391
|
+
wisent/examples/scripts/results/test_medtext_pairs.json,sha256=gCX_ukp2Vz3XHVvXbHesJrMClk-jAUpL_hGiHDPuoRU,1132
|
|
1392
|
+
wisent/examples/scripts/results/test_mela_evaluation.json,sha256=YdoMUr0kc9mRISC4VX2LqBTgvYhwj9lnhFVM0jxDza4,1943
|
|
1393
|
+
wisent/examples/scripts/results/test_mela_pairs.json,sha256=1NcEOC-CypJNCGS2Z25bf4TKEniHE4sLwY0nd8NAMYk,779
|
|
1394
|
+
wisent/examples/scripts/results/test_meqsum_evaluation.json,sha256=T5cQUUNtSigyJKEhnGZokz2u9cTSZWJY8g4_1XRBZMo,1070
|
|
1395
|
+
wisent/examples/scripts/results/test_meqsum_pairs.json,sha256=6iAKdPeN42vvxkbP8JEDcNYN8G6J0o6hYqgpSOzld70,357
|
|
1396
|
+
wisent/examples/scripts/results/test_mercury_evaluation.json,sha256=7__Lb975QG74ongFxumyhuNADJjEK8v4HKxlsfPo-i4,3149
|
|
1397
|
+
wisent/examples/scripts/results/test_mercury_pairs.json,sha256=qGSktBzVt_D5eEbUOw8q7buwpy9QlNynSHskgIf_EMU,1101
|
|
1398
|
+
wisent/examples/scripts/results/test_metabench_evaluation.json,sha256=_Djazx2oyBNDMjmYfobqs--ZS-VK2Y0LizaB8K8b16g,2011
|
|
1399
|
+
wisent/examples/scripts/results/test_metabench_pairs.json,sha256=4f698yHCNn45jWtvsi937vlDwXc_bc8keBR1t7_WwME,1370
|
|
1400
|
+
wisent/examples/scripts/results/test_mgsm_evaluation.json,sha256=FDIgggevW6B6v-LVpG35Adnqevu0rhzqpFLFYuBGgCg,1905
|
|
1401
|
+
wisent/examples/scripts/results/test_mgsm_pairs.json,sha256=awNlZYzMZAnpvqIYAsmYIpX__dNpyULkK8SXB2eWtg0,1247
|
|
1402
|
+
wisent/examples/scripts/results/test_mimic_repsum_evaluation.json,sha256=HEx6y6epZtVORZ5SUcddcW2QaPQ2RWDAVcQrOOUaBBE,8379
|
|
1403
|
+
wisent/examples/scripts/results/test_mimic_repsum_pairs.json,sha256=6xrehIiqBprm0UlbTXcjDiClhkVNR2OCplerG4FbHrE,2679
|
|
1404
|
+
wisent/examples/scripts/results/test_minerva_math_evaluation.json,sha256=C0KqtQDoDVvQUrBOQi71y8EDy5XGjgH8Jh842-699hA,1959
|
|
1405
|
+
wisent/examples/scripts/results/test_minerva_math_pairs.json,sha256=x1tDTGTuf1Yz_M5EHN5c_mQUZH98ZlVRu_mbsXhxEvY,635
|
|
1406
|
+
wisent/examples/scripts/results/test_mlqa_evaluation.json,sha256=DfvQWNrJc6dVzHIaAP6Cu6VafX5E88eCi6Ec2r3_FuI,2261
|
|
1407
|
+
wisent/examples/scripts/results/test_mlqa_pairs.json,sha256=_nGqCBrp3b3AB9jByXnn9EqyJkAMffEaXJrp0vH7LrY,12808
|
|
1408
|
+
wisent/examples/scripts/results/test_mmlu-pro-plus_evaluation.json,sha256=fr6x3pn8JXjm7j3ZTvdiFW__SZCxQJNiSlBgjgRhxqI,1673
|
|
1409
|
+
wisent/examples/scripts/results/test_mmlu-pro-plus_pairs.json,sha256=LYLIMLEPk26hTFGdPa9wC600Yr7BSdIdtFy5eBoJh4s,1281
|
|
1410
|
+
wisent/examples/scripts/results/test_mmlu_evaluation.json,sha256=tuLggxat1t0pZHRyKWc5-gI_doo0CTR7DPigcERrVn8,2224
|
|
1411
|
+
wisent/examples/scripts/results/test_mmlu_pairs.json,sha256=NR0RPFx7gpxLk10bK5JJW-OPEm3sKARkHdx1Au6egB8,1212
|
|
1412
|
+
wisent/examples/scripts/results/test_mmlu_pro_evaluation.json,sha256=3xCCgz2waH5GcaQ8ORKr8Cy6_kXDSz8OLzRNYQTM1oo,1703
|
|
1413
|
+
wisent/examples/scripts/results/test_mmlu_pro_pairs.json,sha256=gRzICeSFw9FmONXL85pDuInxkD5aqdaGygfSFWWNdRU,1476
|
|
1414
|
+
wisent/examples/scripts/results/test_mmlu_prox_evaluation.json,sha256=sk_jcTTRMZrbQLD8mpxrF6xncdC_DiV9FLsakIVAIIU,2536
|
|
1415
|
+
wisent/examples/scripts/results/test_mmlu_prox_pairs.json,sha256=VISvlVzvo7SbVtg9-Hq2XWhfRJyqjYJRxFciIZ6CRWU,2398
|
|
1416
|
+
wisent/examples/scripts/results/test_mmlusr_evaluation.json,sha256=cJKhoFSSdCYc5bVDqC0rVoifhyWlRwXZUu1E_-1GLh8,869
|
|
1417
|
+
wisent/examples/scripts/results/test_mmlusr_pairs.json,sha256=5apvFNyzubhlep2ARl_-g055Hfn-Hq-f7aiO8-a-5Yk,230
|
|
1418
|
+
wisent/examples/scripts/results/test_mmmu_evaluation.json,sha256=4MVDj3KlgJg_LKZvlcJUfDrZoulOog04iJDwpsqS4hg,1547
|
|
1419
|
+
wisent/examples/scripts/results/test_mmmu_pairs.json,sha256=GNM-Pm-LSoLuOAkpg1fC-xAnAHG7Yo_QJ_BE8I53k4w,406
|
|
1420
|
+
wisent/examples/scripts/results/test_mnli_evaluation.json,sha256=PlDAWmGKjIbhxq6t81vUMbXwx6ssLZOBk2w_LaahQ5k,888
|
|
1421
|
+
wisent/examples/scripts/results/test_mnli_pairs.json,sha256=Pap-Hwf-p65DEgSqim6tFbcuwAMXJvds9gB5vkqsT5M,233
|
|
1422
|
+
wisent/examples/scripts/results/test_model_written_evals_evaluation.json,sha256=5rcANb0-W_ckW5OOtzx6US4RqxSSzsYcHT9pSWWTFIg,1603
|
|
1423
|
+
wisent/examples/scripts/results/test_model_written_evals_pairs.json,sha256=8aGjw1tE7oaMBd6ac7_wG0cwHDZPeWbX5Keq9o6Fr7w,1163
|
|
1424
|
+
wisent/examples/scripts/results/test_moral_stories_evaluation.json,sha256=Rf_hHwFRkrKhhtPSN3BmZmCEhWp0Qmiw_3xnikMax9g,1238
|
|
1425
|
+
wisent/examples/scripts/results/test_moral_stories_pairs.json,sha256=pEI0KX8TcmEx8BoVQQTtcA-BKgDpNpuYKL973ice68A,415
|
|
1426
|
+
wisent/examples/scripts/results/test_mts_dialog_evaluation.json,sha256=8zti1z8b_djL1UB01BzNt4nZBfKJO1injNUa3hD9LM8,2544
|
|
1427
|
+
wisent/examples/scripts/results/test_mts_dialog_pairs.json,sha256=fi6mGGIcT_LjNHrRpr9NISMc_VPls6n0trSSXfjQ-No,1243
|
|
1428
|
+
wisent/examples/scripts/results/test_multiblimp_evaluation.json,sha256=Vp-rOHvLtqu051a2rX__hz-TwXiEZc2yqLh6d01t6PA,2641
|
|
1429
|
+
wisent/examples/scripts/results/test_multiblimp_pairs.json,sha256=o7VTB8oeu9E_SJFeW5vUydoyCmNeN50g8opvnCu_rIM,1008
|
|
1430
|
+
wisent/examples/scripts/results/test_multimedqa_evaluation.json,sha256=rY3Hy5wWLd47sESGw3GIUrdO_miKBWeVlkPCGM6MJUo,1661
|
|
1431
|
+
wisent/examples/scripts/results/test_multimedqa_pairs.json,sha256=z-r1cibi5LPUp5-O2C7oL0q-7DXHh7b7sicD2Zn9z0c,1094
|
|
1432
|
+
wisent/examples/scripts/results/test_multipl_e_evaluation.json,sha256=vlfKWPP0dK2QbZz5-uCQMh480At1yyWRkX0AkxcoRqI,4579
|
|
1433
|
+
wisent/examples/scripts/results/test_multipl_e_pairs.json,sha256=MzcZml2RBBD0DlAmP-vs3xZXtoT0m3wD2pGsCe6YukM,1668
|
|
1434
|
+
wisent/examples/scripts/results/test_mutual_evaluation.json,sha256=IGoKKW2kCb5XPm5l08w7Tk2hVc-aA_vBm6aBG3GWrt4,1347
|
|
1435
|
+
wisent/examples/scripts/results/test_mutual_pairs.json,sha256=8-u5J_cx_0IxYKaiJklBEuDreE8FLohc3JARCC4l52Q,1033
|
|
1436
|
+
wisent/examples/scripts/results/test_non_greedy_robustness_agieval_aqua_rat_evaluation.json,sha256=MUr_9WEOuunLgQeh7NSMwQQuRRDWu7WMZiGwKURFAqw,979
|
|
1437
|
+
wisent/examples/scripts/results/test_non_greedy_robustness_agieval_aqua_rat_pairs.json,sha256=hlfqPK1o0KodtMbH4q3nKZFU6j6erXUzz7lOGg_NQuk,514
|
|
1438
|
+
wisent/examples/scripts/results/test_noreval_evaluation.json,sha256=ine0Fxve9kpdESw67UIOwIFlRi8p6FdoLFVjdzvUgRA,1555
|
|
1439
|
+
wisent/examples/scripts/results/test_noreval_pairs.json,sha256=gnPbR9uHRjxueKc1d6zrCVZz6PD1_c_KrZsW0ExV4VU,362
|
|
1440
|
+
wisent/examples/scripts/results/test_noticia_evaluation.json,sha256=v9I8VwXLfjB2TMLf_zERIvmWUNuaNsxoRz75qi_ra6o,1114
|
|
1441
|
+
wisent/examples/scripts/results/test_noticia_pairs.json,sha256=P5O5Tc67RNV45A_DRJgLyPDngzKqWs_X8oxpy4w-wFw,2268
|
|
1442
|
+
wisent/examples/scripts/results/test_nq_open_evaluation.json,sha256=CZZNIoRpjNDbScxC9nxWwSn3v25httU0Zz065zrXNyU,975
|
|
1443
|
+
wisent/examples/scripts/results/test_nq_open_pairs.json,sha256=vmsi1wyYZMVYmxHMdDMGrWTGNk7McE4s9JA8SC8x7BI,258
|
|
1444
|
+
wisent/examples/scripts/results/test_olaph_evaluation.json,sha256=tlFZzxONRbkhJjUHEmOBPo06qx130O-piPVXu5GDnMU,3322
|
|
1445
|
+
wisent/examples/scripts/results/test_olaph_pairs.json,sha256=sLKs-cEEg8wOIcG7bL5-REel2UPfnofgTt2PtBNuydc,921
|
|
1446
|
+
wisent/examples/scripts/results/test_openbookqa_evaluation.json,sha256=YmXM2mR7WE5-7X6H5EwTiedo6QY8o-ozrJIZuQ7rPTg,927
|
|
1447
|
+
wisent/examples/scripts/results/test_openbookqa_pairs.json,sha256=L_VshVYJlou9bj8GfZQPpbgR5ycTVrV75UmMkajzdyU,258
|
|
1448
|
+
wisent/examples/scripts/results/test_openllm_evaluation.json,sha256=RA7y4wESjsBdEoXqBR7SjWVOnK1Ep8Tk7lf-McXYCS4,3589
|
|
1449
|
+
wisent/examples/scripts/results/test_openllm_pairs.json,sha256=gc3r1_eo3BdM_V_oiJ36hVXanyqCcj2AYP7wUc6MjhI,2068
|
|
1450
|
+
wisent/examples/scripts/results/test_option_order_robustness_agieval_aqua_rat_evaluation.json,sha256=JCpnuTyJTqhRNnMaz70-Ksfu31wDjkHOxXTP4Hczhkw,981
|
|
1451
|
+
wisent/examples/scripts/results/test_option_order_robustness_agieval_aqua_rat_pairs.json,sha256=hlfqPK1o0KodtMbH4q3nKZFU6j6erXUzz7lOGg_NQuk,514
|
|
1452
|
+
wisent/examples/scripts/results/test_paloma_evaluation.json,sha256=cTLr0S6bGMkkK6ywp-0YlTzdF2yzHIfUyOyReaAlBuU,13794
|
|
1453
|
+
wisent/examples/scripts/results/test_paloma_pairs.json,sha256=1NCnC87kemagr0Qv1Ec6u5YTsjpQSAy4IOv8QdY911g,4378
|
|
1454
|
+
wisent/examples/scripts/results/test_paws-x_evaluation.json,sha256=ZhxMHD8OqUw1xaM4UMlckSG5lgWJrm3Vm9KhbAXKOOo,1844
|
|
1455
|
+
wisent/examples/scripts/results/test_paws-x_pairs.json,sha256=QomDQVAjgmRiOYAtnuZkYevFjwPyNmLPZDgpYxiAv_k,887
|
|
1456
|
+
wisent/examples/scripts/results/test_penn_treebank_evaluation.json,sha256=whE__BRLDMW-cFr20TJ5za8A4wvLqFxZTBCLlmWaPdw,3722
|
|
1457
|
+
wisent/examples/scripts/results/test_penn_treebank_pairs.json,sha256=_AWjDaXjel4jaiFb1IOk-bZ3wqNi9Yg4e59dZSOYlPc,1565
|
|
1458
|
+
wisent/examples/scripts/results/test_piqa_evaluation.json,sha256=P6YyDegF8wSsyGmuzQsynkzbXA1LCTmBI59vbcwTqEk,1839
|
|
1459
|
+
wisent/examples/scripts/results/test_piqa_pairs.json,sha256=1wH2ylQidQpuy1UvxaBVONjjc7zC8Pa2cT3rvp_RT0Q,849
|
|
1460
|
+
wisent/examples/scripts/results/test_polemo2_evaluation.json,sha256=FOO99tAhAoQwthVaPUw5P9nUk31zkgPB2WHJoVgLKW8,911
|
|
1461
|
+
wisent/examples/scripts/results/test_polemo2_pairs.json,sha256=LjWnibwZ9GWrfZ5kSiNmi3nZXoWmtsEvW_Hcqf-kgvI,429
|
|
1462
|
+
wisent/examples/scripts/results/test_polymath_en_high_evaluation.json,sha256=PTCjnInzvE8JAblMzvmx8m5EudBMqrf74nBX1clNkkg,875
|
|
1463
|
+
wisent/examples/scripts/results/test_polymath_en_high_pairs.json,sha256=vXaNI8UtVGZo7ysrQGWYh47GcumeQF6_VQ4ngRg8zNE,225
|
|
1464
|
+
wisent/examples/scripts/results/test_polymath_en_medium_evaluation.json,sha256=jEX02vrP9l_ticRXaG13GKTxemhRxI00Ens0K88Mca8,979
|
|
1465
|
+
wisent/examples/scripts/results/test_polymath_en_medium_pairs.json,sha256=vjC0qByKtHt4xMFG9OUPu4pUGtLx415CFLZ4MJRjop0,290
|
|
1466
|
+
wisent/examples/scripts/results/test_polymath_zh_high_evaluation.json,sha256=f-z_ZxD12J45n-8Nz0Emno3lwvkuhXIla9M3EvthZjo,935
|
|
1467
|
+
wisent/examples/scripts/results/test_polymath_zh_high_pairs.json,sha256=t8i5ZriVIT_KyyatYwcvJ16NOcHrDxhN41Z_lY2A4V8,270
|
|
1468
|
+
wisent/examples/scripts/results/test_polymath_zh_medium_evaluation.json,sha256=AGapX04mz2KzDEVLtgC4EPq8wuRzEOhoFm-gEyfz7ms,1061
|
|
1469
|
+
wisent/examples/scripts/results/test_polymath_zh_medium_pairs.json,sha256=a2nLV56FBnLPNp3C5gctjoei6iUzLP4JTWJ3mp2CT8k,349
|
|
1470
|
+
wisent/examples/scripts/results/test_portuguese_bench_evaluation.json,sha256=8dsVknTWuTKda85YoVKmhi5YU-sqNoRJu2yXG7HQmgc,4000
|
|
1471
|
+
wisent/examples/scripts/results/test_portuguese_bench_pairs.json,sha256=0pwL4bTLRJ6uXbkIMnLWYoD4JPXqT2Fbcy1G0UG3BNQ,2190
|
|
1472
|
+
wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat_evaluation.json,sha256=JEQ_-iDoK5au-IVf2idaaxs5lLkRQkcfK9O2l0q93_E,975
|
|
1473
|
+
wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat_pairs.json,sha256=hlfqPK1o0KodtMbH4q3nKZFU6j6erXUzz7lOGg_NQuk,514
|
|
1474
|
+
wisent/examples/scripts/results/test_prost_evaluation.json,sha256=wJj51-v8JEpcMvROsItrMBmsV_Uh98KB7TnXQ2Hb4E0,883
|
|
1475
|
+
wisent/examples/scripts/results/test_prost_pairs.json,sha256=0UA-IVEhtaS7AQ_QrFC87qhXTnXSZNye_CQecxdPc2M,253
|
|
1476
|
+
wisent/examples/scripts/results/test_ptb_evaluation.json,sha256=Gz2JHDOGGd7jFdK8y-UgLxPK9CV42Rql0CbLsljXviM,3712
|
|
1477
|
+
wisent/examples/scripts/results/test_ptb_pairs.json,sha256=2XajZW4Db3MDXK4lkPxwwoWFZfQ8Sk3xmpxef36UpBw,1565
|
|
1478
|
+
wisent/examples/scripts/results/test_pubmedqa_evaluation.json,sha256=GM4Lrp6jQ7wqfKr3KC9pkA36XsHb6x8x0ZS1jaUYe4I,869
|
|
1479
|
+
wisent/examples/scripts/results/test_pubmedqa_pairs.json,sha256=krcgQCockT7B4fgDty68L_lqFa8IPSB85JsJROjIZYg,1557
|
|
1480
|
+
wisent/examples/scripts/results/test_pythia_evaluation.json,sha256=B4hBcCinToy7vucnVoH5Y2zrGzHu0-UYpHXkk1_7Yk0,1806
|
|
1481
|
+
wisent/examples/scripts/results/test_pythia_pairs.json,sha256=tm8oBMDIJ0Zb_x-V9H-fjc4MCdfHsK4-7jG8kaYz7og,413
|
|
1482
|
+
wisent/examples/scripts/results/test_qa4mre_evaluation.json,sha256=d91z1t8qpUPmcohmF60rbeVMXH3hVpcWbyWPV94IdHY,1023
|
|
1483
|
+
wisent/examples/scripts/results/test_qa4mre_pairs.json,sha256=8dzioTuq2j3ZeDRoYhKqzMHr1YolLu8bEGFN4nT-E4U,8745
|
|
1484
|
+
wisent/examples/scripts/results/test_qasper_evaluation.json,sha256=ins_h22cFfVkEzqnEUc6gBfzSbieI1QtKKst081Ya7M,1072
|
|
1485
|
+
wisent/examples/scripts/results/test_qasper_pairs.json,sha256=CdR7bklX5esyhk36hH8bvW7WDeBKKKElvnBTvHMwbn4,1477
|
|
1486
|
+
wisent/examples/scripts/results/test_race_evaluation.json,sha256=zCXDua0wj5kFSvgSb_N3KUo5eSQiKTxQugxyRGDjEVM,1049
|
|
1487
|
+
wisent/examples/scripts/results/test_race_pairs.json,sha256=uKxuV24iH5hD87fIe3A1_o9MIHBjapH7EPOIcTVNT3I,1888
|
|
1488
|
+
wisent/examples/scripts/results/test_realtoxicityprompts_evaluation.json,sha256=AliJVbmT2TzflHaNQ54XftN5KZqJvQ4I6wQ-cJFxmlA,1383
|
|
1489
|
+
wisent/examples/scripts/results/test_realtoxicityprompts_pairs.json,sha256=-ZRWtUI0qSAWOx7PP3uXBWtOvq32JsqQdxszgY-16UY,351
|
|
1490
|
+
wisent/examples/scripts/results/test_recode_evaluation.json,sha256=BkHn0HZlhc0P2YM9GB16S-Su-rOuI4VcSsiN3ML47WM,3389
|
|
1491
|
+
wisent/examples/scripts/results/test_recode_pairs.json,sha256=-P3lrMjLrU47kjK1A-DZTGE2oRVZ-UdttRGvblom6aY,1277
|
|
1492
|
+
wisent/examples/scripts/results/test_record_evaluation.json,sha256=cA02Mecr4Ygb050Jk3AjeM5j6_9c1FZ9j52snPqext0,897
|
|
1493
|
+
wisent/examples/scripts/results/test_record_pairs.json,sha256=AFNH1tSovcNwBtJlBKV9fSlVBCw-2FsrWPV4cfgOWwc,1427
|
|
1494
|
+
wisent/examples/scripts/results/test_ruler_evaluation.json,sha256=xo49T3t1QmjQS-k0mZbxSZeV79bOZwl7F_ros4-DN-E,1719
|
|
1495
|
+
wisent/examples/scripts/results/test_ruler_pairs.json,sha256=cBQbgenfSKJfoHKgaMNkfLvw_Ozm1i7WjDyCJ3zp4fk,28484
|
|
1496
|
+
wisent/examples/scripts/results/test_sciq_evaluation.json,sha256=dNqTNtGCGziK3A6b-uoL_Uz_h2K1-Nk1vUDvu8JonRc,909
|
|
1497
|
+
wisent/examples/scripts/results/test_sciq_pairs.json,sha256=CG9VCEWOT0S3gR3RzzUAGi7oUpC2yHoACfbUUsyAmtg,537
|
|
1498
|
+
wisent/examples/scripts/results/test_score_evaluation.json,sha256=j-liQQqy3apfNakCcgnt1I2f-ZUWF0I09KQx2U-lNYQ,1773
|
|
1499
|
+
wisent/examples/scripts/results/test_score_pairs.json,sha256=hoG_hQSfSEPveBILiGzXNjGyW9DOXTWtEl6fDvYKQFw,581
|
|
1500
|
+
wisent/examples/scripts/results/test_self_consistency_evaluation.json,sha256=B9ZoGJlhOIbhuFLt2zNG2rCKxow5tqHIn8U2NyRk864,878
|
|
1501
|
+
wisent/examples/scripts/results/test_self_consistency_pairs.json,sha256=yxW6_k2XylyEqPMPhZhwxyCnw-vwCYBedSkx1o5Zeg4,403
|
|
1502
|
+
wisent/examples/scripts/results/test_siqa_evaluation.json,sha256=1alziLCUyDX_7o0N5sbNndysZBEaYSzR7luyVLnvTnY,987
|
|
1503
|
+
wisent/examples/scripts/results/test_siqa_pairs.json,sha256=fEDTSRyrTHBW1zgfjmP6XhonINqLcwsMnQOGqRF3RDs,347
|
|
1504
|
+
wisent/examples/scripts/results/test_spanish_bench_evaluation.json,sha256=wCfI38yE6Aln-ZjEkQfe2kIYsbjYSwD8UXByG792u00,2663
|
|
1505
|
+
wisent/examples/scripts/results/test_spanish_bench_pairs.json,sha256=7CNMc4P5A5oxc-5PEF3hRvYU3lD8Az2VMzIL-1Zb2NI,1235
|
|
1506
|
+
wisent/examples/scripts/results/test_squad2_evaluation.json,sha256=FuhDnQzcmU_KCjB0XEF8TNLJ21rd-YIR8VHTIw5OvAY,969
|
|
1507
|
+
wisent/examples/scripts/results/test_squad2_pairs.json,sha256=lO_MW2N-KxNymjgIqdraiRx5814j2B5QJ-hGjpXAhmI,1055
|
|
1508
|
+
wisent/examples/scripts/results/test_squadv2_evaluation.json,sha256=Ptp6ftFeIaep_18jUQbnDooiWl_-Rb1oZL2H9HEBWBQ,978
|
|
1509
|
+
wisent/examples/scripts/results/test_squadv2_pairs.json,sha256=9gmSl12mrGAaXL0TDebZO0yFefvSTiduZT3MpUO0Akk,1063
|
|
1510
|
+
wisent/examples/scripts/results/test_super-glue-lm-eval-v1-seq2seq_evaluation.json,sha256=dHD9bopxO_oPFHdwVN-X5VkREK7_ba3XTU5s1E_ZkkM,888
|
|
1511
|
+
wisent/examples/scripts/results/test_super-glue-lm-eval-v1-seq2seq_pairs.json,sha256=YZiJJSLsFTIRMfctwiXSEStI3YaznaWTbF58-eGE8OQ,1583
|
|
1512
|
+
wisent/examples/scripts/results/test_super-glue-lm-eval-v1_evaluation.json,sha256=ETuEkyK8GwV09yd9OE8KmjWyZzpRxGnFvM9fWPHmBrE,1654
|
|
1513
|
+
wisent/examples/scripts/results/test_super-glue-lm-eval-v1_pairs.json,sha256=ASGZcutkQv-dqflNqC70xTf8c0vfqC-cCo8FO_A_VaA,2729
|
|
1514
|
+
wisent/examples/scripts/results/test_swag_evaluation.json,sha256=WbIntRdAMsRkgkVkhXH495RnEfLd8flMNVZd1XA_aiw,1043
|
|
1515
|
+
wisent/examples/scripts/results/test_swag_pairs.json,sha256=fj-KXPXdS4e2Lt8lTGYz7QkOxvQ5yDpfoOrA9CiYYys,288
|
|
1516
|
+
wisent/examples/scripts/results/test_tinyBenchmarks_evaluation.json,sha256=GCN4KDa9HwZctfgBdF_GWXD0-1aEGhh_LK3adzcX7II,1816
|
|
1517
|
+
wisent/examples/scripts/results/test_tinyBenchmarks_pairs.json,sha256=KlTcdNMmXAlwArOmV3qLdfjTTSygO8Elu_eZWK3k-5A,1086
|
|
1518
|
+
wisent/examples/scripts/results/test_tmmluplus_evaluation.json,sha256=epDgRWk1EOr0x3EyNP8GFPMa5-L6teec1VaW-b0dy6g,3438
|
|
1519
|
+
wisent/examples/scripts/results/test_tmmluplus_pairs.json,sha256=qiV2NvDXx1h6uvXzldM3vTL_3CR9BbmxmkViXtlw1n8,1590
|
|
1520
|
+
wisent/examples/scripts/results/test_translation_evaluation.json,sha256=ZHJimHvMjVMQCuAlWF2WSG2cV3LdVWLJzgqTldW6WTA,2483
|
|
1521
|
+
wisent/examples/scripts/results/test_translation_pairs.json,sha256=ET1sAjTY8owKBheIr_DHmGLKFb_3hQeIiHKrtpL8RoQ,798
|
|
1522
|
+
wisent/examples/scripts/results/test_triviaqa_evaluation.json,sha256=loiHb2Y7Bl08egPzV2vSLjXFKy_VpVKUEmNPdGjZXpI,923
|
|
1523
|
+
wisent/examples/scripts/results/test_triviaqa_pairs.json,sha256=Pry_8uSNJ43u6pJvBI8AHPsWkCMeHO_s8lqy7u7ilVY,222
|
|
1524
|
+
wisent/examples/scripts/results/test_truthfulqa-multi_evaluation.json,sha256=LREOfE9bp4OHZF8dMIyvHbyxCJCVGMp0guB7H6zF62o,2254
|
|
1525
|
+
wisent/examples/scripts/results/test_truthfulqa-multi_pairs.json,sha256=vE5O-aBu9t1_M6JGxEvKIs4y7zxjUaRw0xofP-PzYxI,597
|
|
1526
|
+
wisent/examples/scripts/results/test_truthfulqa_evaluation.json,sha256=B-Fjz1DmOb0SEw0ufyk-HmMIubojb09ArtNCohqK0BM,1374
|
|
1527
|
+
wisent/examples/scripts/results/test_truthfulqa_mc1_evaluation.json,sha256=BqBYV3CP7SPDGz_ZRJ95m_R0lNckDEA4dnyZ5Xgi4pI,1378
|
|
1528
|
+
wisent/examples/scripts/results/test_truthfulqa_mc1_pairs.json,sha256=9NXU1laojUaEQXtoKQDYEImMosgO4MF4dPtdy9dEpgs,557
|
|
1529
|
+
wisent/examples/scripts/results/test_truthfulqa_mc2_evaluation.json,sha256=AeMsc60T14qwt7TJDIpbJqzVO40poR9SWUZfPpMsxHc,1378
|
|
1530
|
+
wisent/examples/scripts/results/test_truthfulqa_mc2_pairs.json,sha256=9NXU1laojUaEQXtoKQDYEImMosgO4MF4dPtdy9dEpgs,557
|
|
1531
|
+
wisent/examples/scripts/results/test_truthfulqa_pairs.json,sha256=9NXU1laojUaEQXtoKQDYEImMosgO4MF4dPtdy9dEpgs,557
|
|
1532
|
+
wisent/examples/scripts/results/test_turkishmmlu_evaluation.json,sha256=AnkNpn32HNhA9ncKMpmhYkfTKbm5j5sPejnZEFBB-y8,2010
|
|
1533
|
+
wisent/examples/scripts/results/test_turkishmmlu_pairs.json,sha256=uZxUHPKY43RsghFDNvSRd6IL9xZUYa3EjvF8C12oei4,934
|
|
1534
|
+
wisent/examples/scripts/results/test_unfair_tos_evaluation.json,sha256=mjW_7wMPOJ4O-arUf16VbodEu3ogzfOBKknCV1AYkLg,918
|
|
1535
|
+
wisent/examples/scripts/results/test_unfair_tos_pairs.json,sha256=fkqUlW6y3QYhWyrQCwYAtOToFoGQjP2G2V73GxQnkkk,488
|
|
1536
|
+
wisent/examples/scripts/results/test_unscramble_evaluation.json,sha256=bu7yzHEYR_xO64FGVb8hkPGp6YYIp1xIOGY6NPoOKCE,1485
|
|
1537
|
+
wisent/examples/scripts/results/test_unscramble_pairs.json,sha256=CgDUkWIMU-TvIE3EHrN0t2bWdcGwNGirfwIDlQXXwGI,262
|
|
1538
|
+
wisent/examples/scripts/results/test_webqs_evaluation.json,sha256=Bg35qWgfylN3Pf0CKvTSMawPcqLyZ10YYJCaEVqnvs8,1045
|
|
1539
|
+
wisent/examples/scripts/results/test_webqs_pairs.json,sha256=ZTXifP3ZwsbTW4F1nRNvTPB0bsJzacXiYqgOCgyYMuI,293
|
|
1540
|
+
wisent/examples/scripts/results/test_wikitext103_evaluation.json,sha256=MhNmXu3J_K-g-faNt1w2wLsEcfPZkmPScXVcc6h3EwI,3720
|
|
1541
|
+
wisent/examples/scripts/results/test_wikitext103_pairs.json,sha256=Lt1hRlaLTqi6S2_YMMhfVjGFTzhzmbeLwE1-UdEAcSo,1565
|
|
1542
|
+
wisent/examples/scripts/results/test_wikitext_evaluation.json,sha256=dzK8PjAo75A56GEtzOrfqpICxd1YIF-lR6L0Kh3ocBo,3717
|
|
1543
|
+
wisent/examples/scripts/results/test_wikitext_pairs.json,sha256=cq8HOwPiu4i5Qjn7WQkUkHTd6Xx-sVzZFz-n_qSvxW0,1565
|
|
1544
|
+
wisent/examples/scripts/results/test_winogender_evaluation.json,sha256=OPiDUAklkfkiw-U9uNeZG69hN3u5Wgs22JcuzAQyK80,1652
|
|
1545
|
+
wisent/examples/scripts/results/test_winogender_pairs.json,sha256=Wr4mgPnv98ZQJNLw7caQyqakL1BSaCB1Dj6an3mEkN4,485
|
|
1546
|
+
wisent/examples/scripts/results/test_winogrande_evaluation.json,sha256=TC2jspGQjHTYLR5kb4ZIr6u3MYssG2E5YhFxDdrjwVg,886
|
|
1547
|
+
wisent/examples/scripts/results/test_winogrande_pairs.json,sha256=pLNpvtDn9398l4JqzwlxdJAFTp1PpYRQSi5WP1oOBcc,233
|
|
1548
|
+
wisent/examples/scripts/results/test_wmdp_evaluation.json,sha256=i21siUjmeY7AbnaM5S1Y7EYuhC8e8uGFuGyDVbexN_s,1106
|
|
1549
|
+
wisent/examples/scripts/results/test_wmdp_pairs.json,sha256=NxZfI9iD6CV4jS3LV7AhoogEZpqnSwTCxfG4p0VPYJA,401
|
|
1550
|
+
wisent/examples/scripts/results/test_wmt-ro-en-t5-prompt_evaluation.json,sha256=bbnszaIIIiCcP6NQ9g28rhDTWTVl0kjZot7_o5QmODg,1168
|
|
1551
|
+
wisent/examples/scripts/results/test_wmt-ro-en-t5-prompt_pairs.json,sha256=iED7CMQMDdWDOlxsh07PyE4YWFqmO-3SoZ8ExzVdXKI,328
|
|
1552
|
+
wisent/examples/scripts/results/test_wmt14_en_fr_evaluation.json,sha256=FMQdeqh5akNJ5sOJ6YxgE1uB_OGhiWBwLlJYq_oy1I4,1147
|
|
1553
|
+
wisent/examples/scripts/results/test_wmt14_en_fr_pairs.json,sha256=ypwRrogw8zm_M10OWOZgggcGCaskcbEYN9-B5SaOMX0,298
|
|
1554
|
+
wisent/examples/scripts/results/test_wmt16_en_de_evaluation.json,sha256=oT86Ix1U9DuULsDgbGrSGrgn63X-Vr4KVr_SxsLRops,941
|
|
1555
|
+
wisent/examples/scripts/results/test_wmt16_en_de_pairs.json,sha256=lmYdCNDTGztU0gCIsVgKZ0yErjdLH-xsqIIkKaycJHg,197
|
|
1556
|
+
wisent/examples/scripts/results/test_wmt16_ro_en_evaluation.json,sha256=9ClcOUtO0vfHn_Zd29tqNUVEFxbRpSyfbXA1sQNmGYY,1131
|
|
1557
|
+
wisent/examples/scripts/results/test_wmt16_ro_en_pairs.json,sha256=hO0i_AfnErQ8ChKyWgiCws0dfY-E07txotfOaJY5ZVs,301
|
|
1558
|
+
wisent/examples/scripts/results/test_wsc273_evaluation.json,sha256=RV9FqeePrCsVYedlAnXksfGEppFtlBzL_P_YIuP_SMg,962
|
|
1559
|
+
wisent/examples/scripts/results/test_wsc273_pairs.json,sha256=yK8EMyi5fR_lx8wb2SdW7Co88tBD8ORJIdeJd4CpPkU,274
|
|
1560
|
+
wisent/examples/scripts/results/test_xcopa_evaluation.json,sha256=40LnQrNzZUabyvWXVyc4JSXDAY6TGeade_rg9cJwZ_o,1893
|
|
1561
|
+
wisent/examples/scripts/results/test_xcopa_pairs.json,sha256=RmrzF37PWhnw5sdzh-yKxBHCjgliNKw2CJ4-JKgYgpo,490
|
|
1562
|
+
wisent/examples/scripts/results/test_xnli_eu_evaluation.json,sha256=_PmCiZDR2xcImFgffTeK6mdHtQrIhq_3xTUcej-mKrg,906
|
|
1563
|
+
wisent/examples/scripts/results/test_xnli_eu_pairs.json,sha256=OnNCubMX4U5Haz3HgXRTEOgnkhkRqWnhlBL3TQJjmFk,302
|
|
1564
|
+
wisent/examples/scripts/results/test_xnli_evaluation.json,sha256=g_ojPO_A6CqpEWey4NVXZ5e13VcverS4jUURAaKZk_U,1843
|
|
1565
|
+
wisent/examples/scripts/results/test_xnli_pairs.json,sha256=lHgJF38o3VUeM9ApvV-5v-j2MobCP8mnFYtEdq7Vpi0,1171
|
|
1566
|
+
wisent/examples/scripts/results/test_xquad_evaluation.json,sha256=FWIrqhB6E9zIHyg0M9UprDedcYg0rNpVsFmpAzYzmsA,2171
|
|
1567
|
+
wisent/examples/scripts/results/test_xquad_pairs.json,sha256=syPySOQjRWsaHEfiWhrLEDOjvV-dCNjcp3uyKqIU86Q,2070
|
|
1568
|
+
wisent/examples/scripts/results/test_xstorycloze_evaluation.json,sha256=3LyG1hQyhyiWLqtgwBT4ZknTZaXililXxtSEwcopx-M,4641
|
|
1569
|
+
wisent/examples/scripts/results/test_xstorycloze_pairs.json,sha256=TOSmSQpSQNB1dklWDENQQkIQ-QtVsg7VFqn0P7AcVnI,2952
|
|
1570
|
+
wisent/examples/scripts/results/test_xsum_evaluation.json,sha256=ahpGvrimsFt1zCkY6iimuBPTSx0OW-0lbRUP5v-OkI8,1419
|
|
1571
|
+
wisent/examples/scripts/results/test_xsum_pairs.json,sha256=goOLYN9jYo-JYIigxdLEsWqgc2EsODU-YeDDWvCv6aE,840
|
|
1572
|
+
wisent/examples/scripts/results/test_xwinograd_evaluation.json,sha256=JyroLtcJj5oeChul7WdadkZsvOAVT18zuEIl3rNFqy0,2210
|
|
1573
|
+
wisent/examples/scripts/results/test_xwinograd_pairs.json,sha256=SDObE7ToPGGKjDrB20WVnCT8BCclvvFk_133Fdv2Img,835
|
|
1574
|
+
wisent/examples/scripts/results/test_yahoo_answers_topics_evaluation.json,sha256=g86lFm2e0ALaJruyKOabK0mkYJHTc1GiYpbpxGuVJqk,923
|
|
1575
|
+
wisent/examples/scripts/results/test_yahoo_answers_topics_pairs.json,sha256=g88bnzqG_95nONXy11ZnNQA-2x_m_R9FbwTtIUMgBgI,846
|
|
1576
|
+
wisent/examples/scripts/results/test_AraDiCE_boolq_egy/test_AraDiCE_boolq_egy_evaluation.json,sha256=8Fif1e-Z5Yj_MXcrLTvrdRv_1_TDGf4F-CcD9P6vc6I,1114
|
|
1577
|
+
wisent/examples/scripts/results/test_AraDiCE_boolq_egy/test_AraDiCE_boolq_egy_pairs.json,sha256=ACvi5qvSFg8N3zmaZ0lp3KFAVO0eIo8mulILuUDA09Q,402
|
|
1578
|
+
wisent/examples/scripts/results/test_aradice/test_aradice_evaluation.json,sha256=DLwvsG6RZmsSFdwvr9rcR7W3jPx5m0aa9vSl6UDF74Y,2195
|
|
1579
|
+
wisent/examples/scripts/results/test_aradice/test_aradice_pairs.json,sha256=R6Gd163_VDKHL933ktNjT630al_JtWzczLWgCWAikGQ,998
|
|
1580
|
+
wisent/examples/scripts/results/test_aradice3/test_aradice_evaluation.json,sha256=pr8FAerYR2UPI6nqbnHVPZmkCGArVHTOE79EcaNdXKY,1948
|
|
1581
|
+
wisent/examples/scripts/results/test_aradice3/test_aradice_pairs.json,sha256=7bV0mBa4IuyOOU0AOLa5GaDVSIm-v6hrMN5SKUpBqA0,751
|
|
1582
|
+
wisent/examples/scripts/results/test_basque/test_basque-glue_pairs.json,sha256=eiN94jLrpFyOEQPSJeu8CXttaLJFCGnnmixiOGRcxEk,570
|
|
1583
|
+
wisent/examples/scripts/results/test_basque2/test_basque-glue_evaluation.json,sha256=Np2XXbtbzl0d9aAcS7rrc6U3yMUE8S4SakKaxFCqQKY,1797
|
|
1584
|
+
wisent/examples/scripts/results/test_basque2/test_basque-glue_pairs.json,sha256=eiN94jLrpFyOEQPSJeu8CXttaLJFCGnnmixiOGRcxEk,570
|
|
1585
|
+
wisent/examples/scripts/results/test_basque_glue/test_basque-glue_evaluation.json,sha256=pMnny7J-O7FWKUZfY7z4fV4BSR2j03TuNpb7B53YEKk,1701
|
|
1586
|
+
wisent/examples/scripts/results/test_basque_glue/test_basque-glue_pairs.json,sha256=Iwt6Kmg3fiWhQXOkOlXtfKksYwsXvQWKV4S_NiEkleM,613
|
|
1587
|
+
wisent/examples/scripts/results/test_boolq/test_boolq_evaluation.json,sha256=b4iXD4lHIrUZjntmXJSXQ-ONFpsdQ_nh7zLSiev-HlA,864
|
|
1588
|
+
wisent/examples/scripts/results/test_boolq/test_boolq_pairs.json,sha256=jK_jGZrMSaJVwiBYYmG9FxYaQhpn4-J9HH5lflr0lZ4,1554
|
|
1589
|
+
wisent/examples/scripts/results/test_ceval/test_ceval_evaluation.json,sha256=wcjDYFj5B8R7UcHxLMLFQR-7w97eIk0nRVl3lf1tPuc,2283
|
|
1590
|
+
wisent/examples/scripts/results/test_ceval/test_ceval_pairs.json,sha256=dk4jGjqCE_HZLAWz6lIivYWRAt6rz0cao2GhhElzGVg,872
|
|
1591
|
+
wisent/examples/scripts/results/test_ceval_accountant/test_ceval-valid_accountant_evaluation.json,sha256=NwZBvgrCpm2ML57l_S2ezkBuD28Hbaq9-BKDCN8GT3A,2339
|
|
1592
|
+
wisent/examples/scripts/results/test_ceval_accountant/test_ceval-valid_accountant_pairs.json,sha256=c0fWeirDrdsrsN2LTUftkcjlMSqhFZig1aZlbokQWSg,938
|
|
1593
|
+
wisent/examples/scripts/results/test_ceval_valid/test_ceval_valid_evaluation.json,sha256=Mn-o6UK1MWuH3EDi1XsSlnuy0wOXMRWk7AUS_NYiE0E,2289
|
|
1594
|
+
wisent/examples/scripts/results/test_ceval_valid/test_ceval_valid_pairs.json,sha256=dk4jGjqCE_HZLAWz6lIivYWRAt6rz0cao2GhhElzGVg,872
|
|
1595
|
+
wisent/examples/scripts/results/test_darija_bench/test_darija_bench_evaluation.json,sha256=mE-UlpD4_bpnlQt-rYW-4S4hS_HXOKpjup-f4pj20hw,2599
|
|
1596
|
+
wisent/examples/scripts/results/test_darija_bench/test_darija_bench_pairs.json,sha256=td8oZ9RFS5qeVx3gNdf6bgo3yjWGd6Tkz7jOa-8zByg,1711
|
|
1597
|
+
wisent/examples/scripts/results/test_eus_exams/test_eus_exams_evaluation.json,sha256=aMWHOmI1qOEPnUadmadu28pYLb4u-_Kxj2OHfHgXs4w,4109
|
|
1598
|
+
wisent/examples/scripts/results/test_eus_exams/test_eus_exams_pairs.json,sha256=Bn3Xm1qJseSwhdAT8j3F4vu3u3K6e3smdMimdorX_JM,2025
|
|
1599
|
+
wisent/examples/scripts/results/test_evalita_llm/test_evalita_llm_evaluation.json,sha256=mukAfoY6hd2R5GiobTbjJavTvMP3tnMbXjIUgahYBHg,1905
|
|
1600
|
+
wisent/examples/scripts/results/test_evalita_llm/test_evalita_llm_pairs.json,sha256=HjC70ItXgdB5iE0LyEILkNb7JHsOrFsb3AGNhZc44zs,1785
|
|
1601
|
+
wisent/examples/scripts/results/test_evalita_mp/test_evalita-mp_te_prompt-1_evaluation.json,sha256=MOI3IDN3T5nThGnfryDslZ1ojyRFyPV6IAnbNkhskHM,889
|
|
1602
|
+
wisent/examples/scripts/results/test_evalita_mp/test_evalita-mp_te_prompt-1_pairs.json,sha256=aTLk2rUAIVnFYSqHhbj2WmLM144_i7u5sgooc3GPd7Y,300
|
|
1603
|
+
wisent/examples/scripts/results/test_evalita_mp2/test_evalita_mp_evaluation.json,sha256=6R1jAtX9sNhfYe9nVQvx4IvGLD4br8klbS_cXMdUNbs,1904
|
|
1604
|
+
wisent/examples/scripts/results/test_evalita_mp2/test_evalita_mp_pairs.json,sha256=HjC70ItXgdB5iE0LyEILkNb7JHsOrFsb3AGNhZc44zs,1785
|
|
1605
|
+
wisent/examples/scripts/results/test_evalita_sp2/test_evalita-sp_sum_task_fp-small_p1_evaluation.json,sha256=7pUXf4PqJLPFOEezImpguOlyQhvTn-WdPEzte3RSmXo,2869
|
|
1606
|
+
wisent/examples/scripts/results/test_evalita_sp2/test_evalita-sp_sum_task_fp-small_p1_pairs.json,sha256=yZ6BL9RAtKIpS0MiWcmiIcq-V3FBuDdkfB-TiU1xhto,3268
|
|
1607
|
+
wisent/examples/scripts/results/test_fld/test_fld_evaluation.json,sha256=S1F5EGJdHSkXIHZxz0YGqWmi_YgK_VJL2UpGHenu_fI,838
|
|
1608
|
+
wisent/examples/scripts/results/test_fld/test_fld_pairs.json,sha256=1L5jg_v5WHopxNJYpEplZXRowQ4hH3bKT3G2TMnW-CQ,1629
|
|
1609
|
+
wisent/examples/scripts/results/test_fld_fixed/test_fld_evaluation.json,sha256=aiEXO4lqEdjtTzY4hFbFTH2CPZXeDPfKfme87WacMhE,842
|
|
1610
|
+
wisent/examples/scripts/results/test_fld_fixed/test_fld_pairs.json,sha256=1L5jg_v5WHopxNJYpEplZXRowQ4hH3bKT3G2TMnW-CQ,1629
|
|
1611
|
+
wisent/examples/scripts/results/test_instruct_humaneval/test_instruct_humaneval_evaluation.json,sha256=ouYU9tVa0qtqLmMj3Z8PdXqh5S4zbFIg8rEekb7cuMU,4640
|
|
1612
|
+
wisent/examples/scripts/results/test_instruct_humaneval/test_instruct_humaneval_pairs.json,sha256=8nPVGlUHfsFdcoRV2Zq30zH9-9dXvrOXXHOJ7yMbvHc,1766
|
|
1613
|
+
wisent/examples/scripts/results/test_inverse_scaling_mc/test_inverse_scaling_mc_evaluation.json,sha256=a9w23kEXr9rzK1T9J_ka9uzjGZlXR5xtocclSs_gNkA,1590
|
|
1614
|
+
wisent/examples/scripts/results/test_inverse_scaling_mc/test_inverse_scaling_mc_pairs.json,sha256=TnxvMiVzO9Df42EZ1VivM1UtwoLtpisMTDR9SNJ5cpQ,1964
|
|
1615
|
+
wisent/examples/scripts/results/test_iwslt2017_ar_en/test_iwslt2017-ar-en_evaluation.json,sha256=lWTEYCVVtqSYGZGEGXZvejR_f2kqfwpOKap0DtCjdds,2307
|
|
1616
|
+
wisent/examples/scripts/results/test_iwslt2017_ar_en/test_iwslt2017-ar-en_pairs.json,sha256=BUu5vtNr-oHb201ONyrDANXSQLNB5XJ6bQ310cyaAJo,1414
|
|
1617
|
+
wisent/examples/scripts/results/test_iwslt2017_en_ar/test_iwslt2017-en-ar_evaluation.json,sha256=qk1QpcJgKhBukSbIjZdy6Xup6XM5YVpSCoIcD8EQRBA,2307
|
|
1618
|
+
wisent/examples/scripts/results/test_iwslt2017_en_ar/test_iwslt2017-en-ar_pairs.json,sha256=qQahZyr7GnOpbfy8RoAf2xAFszGzKzqE_2ZphoBL6Pc,1414
|
|
1619
|
+
wisent/examples/scripts/results/test_iwslt2017_group/test_iwslt2017_evaluation.json,sha256=qyt01KdxO5AfMYIAR1-Cv0l-Ts2u3NEB49t8WOsllt8,2301
|
|
1620
|
+
wisent/examples/scripts/results/test_iwslt2017_group/test_iwslt2017_pairs.json,sha256=Lb8LJ6ykOx6iXOcgcwfSRL4DANGNRg9IMOUfDunM9-0,1414
|
|
1621
|
+
wisent/examples/scripts/results/test_jsonschema_bench/test_jsonschema_bench_evaluation.json,sha256=Vnp1Ha2PyH-VA3IasJtKQ4w9AJKSZldqcV9voMAP980,893
|
|
1622
|
+
wisent/examples/scripts/results/test_jsonschema_bench/test_jsonschema_bench_pairs.json,sha256=rTNcw8zRB8sino6WuO6P2lhKEiRG2VHXSh-a2mhJMdc,19480
|
|
1623
|
+
wisent/examples/scripts/results/test_jsonschema_bench_final/test_jsonschema_bench_evaluation.json,sha256=Vnp1Ha2PyH-VA3IasJtKQ4w9AJKSZldqcV9voMAP980,893
|
|
1624
|
+
wisent/examples/scripts/results/test_jsonschema_bench_final/test_jsonschema_bench_pairs.json,sha256=rTNcw8zRB8sino6WuO6P2lhKEiRG2VHXSh-a2mhJMdc,19480
|
|
1625
|
+
wisent/examples/scripts/results/test_kbl_fixed/test_kbl_evaluation.json,sha256=8kGpFhxjO8x_HRSvAg6-WbTez2E8A51ES0i1b0MOCCM,1980
|
|
1626
|
+
wisent/examples/scripts/results/test_kbl_fixed/test_kbl_pairs.json,sha256=LV-Ua4yLlHsewRWlSYsZwTQsd9Gph9iJlDyrgPTTGdo,1803
|
|
1627
|
+
wisent/examples/scripts/results/test_kormedmcqa/test_kormedmcqa_evaluation.json,sha256=j85PbuvVbOUZCiROCx5s6xGeaDEqGvJmY-zFzC-To-Y,1221
|
|
1628
|
+
wisent/examples/scripts/results/test_kormedmcqa/test_kormedmcqa_pairs.json,sha256=1zItWyoLhYMnSx5dEUSm62Wl47PTgM7IFA0vabZ7yNU,1133
|
|
1629
|
+
wisent/examples/scripts/results/test_kormedmcqa_dentist/test_kormedmcqa_dentist_evaluation.json,sha256=k_oerHtwo5BP0n4tmvk9FY-OqkYPHp-Qrs0K33IAn9I,1135
|
|
1630
|
+
wisent/examples/scripts/results/test_kormedmcqa_dentist/test_kormedmcqa_dentist_pairs.json,sha256=hcU7f-g0h20OCCvcVADoQ7Ri0i6gLl4zPRG4QvLSL9E,837
|
|
1631
|
+
wisent/examples/scripts/results/test_lambada_final/test_lambada_openai_mt_stablelm_en_evaluation.json,sha256=y3Cm3YbjBv_RIXT2FTgaEEADBu9stGdGUgWMpkQ0SXk,903
|
|
1632
|
+
wisent/examples/scripts/results/test_lambada_final/test_lambada_openai_mt_stablelm_en_pairs.json,sha256=flzZN9GO7iUf6V22Mfqtyuktrdd3rK5ReKRNFwNIA4Q,452
|
|
1633
|
+
wisent/examples/scripts/results/test_lambada_multilingual/test_lambada_multilingual_evaluation.json,sha256=2007-h4hL1i6TQwvc5Afu8XZYOZlJN7unUEmJSqXph8,1644
|
|
1634
|
+
wisent/examples/scripts/results/test_lambada_multilingual/test_lambada_multilingual_pairs.json,sha256=vcJdBqSpnZ8kEzehkctGhi_LNsM66X3G1C_44hRJaiU,916
|
|
1635
|
+
wisent/examples/scripts/results/test_lambada_stablelm_en_fixed/test_lambada_openai_mt_stablelm_en_evaluation.json,sha256=TxQIkEKMuXBfr8CFIxDw-bOE3_AfQN6DA96oPs3lfcs,901
|
|
1636
|
+
wisent/examples/scripts/results/test_lambada_stablelm_en_fixed/test_lambada_openai_mt_stablelm_en_pairs.json,sha256=LyMRob8O1u3P3Z_5scO0_vCPIWN7YFyCJqF022Z5GRs,451
|
|
1637
|
+
wisent/examples/scripts/results/test_lambada_stablelm_fixed/test_lambada_openai_mt_stablelm_en_evaluation.json,sha256=bqWXs_-P5t3N_iHZ_8fuLkD28qjypiwDF_CApN26-Zo,903
|
|
1638
|
+
wisent/examples/scripts/results/test_lambada_stablelm_fixed/test_lambada_openai_mt_stablelm_en_pairs.json,sha256=1z3YrPxsSBnpotKl-xbzG3M7Ot4d2hCXVR4FkC-zSUI,452
|
|
1639
|
+
wisent/examples/scripts/results/test_libra/test_libra_evaluation.json,sha256=aVk-MyKvwga9sxUUgnG6DOSaA2xlO2vBAlG6jxgZv_Y,2485
|
|
1640
|
+
wisent/examples/scripts/results/test_libra/test_libra_pairs.json,sha256=v-b8smewd_4we1vsxOkyIn0Ol_oCi5xPv_5conndrgk,880379
|
|
1641
|
+
wisent/examples/scripts/results/test_mastermind/test_mastermind_evaluation.json,sha256=f69woRnb2_yIiKm6JNTpM17z3YkisTqiYgwTgvv8Fzc,1846
|
|
1642
|
+
wisent/examples/scripts/results/test_mastermind/test_mastermind_pairs.json,sha256=nZiJ_cPmiPfl799XB7t5vejrct1fZYCG_dWvHiGbKS0,1778
|
|
1643
|
+
wisent/examples/scripts/results/test_mastermind_24_easy/test_mastermind_24_easy_evaluation.json,sha256=E29v402NEemeza5RfRX-2Vpqykc4eBgKstGpF3wWN0E,935
|
|
1644
|
+
wisent/examples/scripts/results/test_mastermind_24_easy/test_mastermind_24_easy_pairs.json,sha256=x7HZA2lND47FIAVS6XuqfenqPYpv0Vi5z2e293V9jiM,531
|
|
1645
|
+
wisent/examples/scripts/results/test_med_concepts_qa/test_med_concepts_qa_evaluation.json,sha256=ouFMxeZaFneONKOFiCabnHUhi6GK5nqN0v0BLYwj_W4,1577
|
|
1646
|
+
wisent/examples/scripts/results/test_med_concepts_qa/test_med_concepts_qa_pairs.json,sha256=QRn1FPe4dRcoE1E1DzIlg7hCqnVdeBZiXSeteJW44Kk,926
|
|
1647
|
+
wisent/examples/scripts/results/test_med_concepts_qa_atc_easy/test_med_concepts_qa_atc_easy_evaluation.json,sha256=Ozg2aC9NwtF173Tc9TWKXcJAv0EiN2QaxFNDXweTgQs,877
|
|
1648
|
+
wisent/examples/scripts/results/test_med_concepts_qa_atc_easy/test_med_concepts_qa_atc_easy_pairs.json,sha256=RtJxjGHpn-gCGedHZZ5sM2bagfSFyGVF4Vf2z9dhedw,333
|
|
1649
|
+
wisent/examples/scripts/results/test_meddialog_raw_perplexity/test_meddialog_raw_perplexity_evaluation.json,sha256=Ffg25sLagyLYRyefmps_P8-stOfF-FttPiCx84qQaT8,3401
|
|
1650
|
+
wisent/examples/scripts/results/test_meddialog_raw_perplexity/test_meddialog_raw_perplexity_pairs.json,sha256=XneZGQh21Uy9Z0w3W7SHik7Wt7Tt1LzlXzXGAZdkpGQ,1124
|
|
1651
|
+
wisent/examples/scripts/results/test_passkey/test_passkey_evaluation.json,sha256=qKB-mMP9W933gVTJk55hPYxyfDUZFbuYR6SzF2hhVGc,1303
|
|
1652
|
+
wisent/examples/scripts/results/test_passkey/test_passkey_pairs.json,sha256=9lC5EuXXTw_wcsfViJC7xnmojlUBkBDy4U6NMY35Wm8,38658
|
|
1653
|
+
wisent/examples/scripts/results/test_paws_en/test_paws_en_evaluation.json,sha256=zF4GHcRrj7SA3iDc546R0aNlrJXN7_iUKO1lpFndnqk,868
|
|
1654
|
+
wisent/examples/scripts/results/test_paws_en/test_paws_en_pairs.json,sha256=3AMSPsHmYHKiO8hQVAfkcWICtsa3Sq731P6jYXO8YaU,369
|
|
1655
|
+
wisent/examples/scripts/results/test_pile_10k/test_pile_10k_evaluation.json,sha256=iif8Qd0STfl684SafDCF81InL6-FuaF6T63Kh96JjKw,3581
|
|
1656
|
+
wisent/examples/scripts/results/test_pile_10k/test_pile_10k_pairs.json,sha256=4HmOQNqCczfI6OEZeld1HU3HydzT06qD_68QIXJNiPM,1544
|
|
1657
|
+
wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat/test_prompt_robustness_agieval_aqua_rat_evaluation.json,sha256=v_6MzluhERLu25Tc7Lml6kXsvddV6MCLVBm49VBcR2s,975
|
|
1658
|
+
wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat/test_prompt_robustness_agieval_aqua_rat_pairs.json,sha256=cdRa2UXD_Zgta_OSSpHr36knjDOxf7sgTAcZzDK6WI0,570
|
|
1659
|
+
wisent/examples/scripts/results/test_siqa/test_siqa_evaluation.json,sha256=mlGNX3VgWyXk4tBBcB3c7oQ87LYppF4ohlSRDN5pkw4,959
|
|
1660
|
+
wisent/examples/scripts/results/test_siqa/test_siqa_pairs.json,sha256=8xrVmzRdfITwngaXf5J78WJKq_WEPPiRQLpYES0k2HE,294
|
|
1661
|
+
wisent/parameters/__init__.py,sha256=dwyW8F6LSMPM-H4zQX6gn7x5qPs78oNAMYRiRPJ8UCk,36
|
|
1662
|
+
wisent/parameters/lm_eval/all_lm_eval_task_families.json,sha256=MAFcazf3UoDplgaeWdkcKpAOzFRHyc5yLoXLqOwyFvc,2555
|
|
1663
|
+
wisent/parameters/lm_eval/broken_in_lm_eval.json,sha256=C8vWqqJBRSUnLwCnuxwt_hfpaykfkxXN8NRJUCUixzs,120
|
|
1664
|
+
wisent/parameters/lm_eval/evaluations_not_lm_eval_tasks.json,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1665
|
+
wisent/parameters/lm_eval/evaluator_check.json,sha256=XrfAOzyw3SYSIfqOo2Q_U-nXpM13ck4m-aRtL6Za9cg,131833
|
|
1666
|
+
wisent/parameters/lm_eval/final_verification.json,sha256=clscpavGxNXbX2Xn-idHKXGO0_fhMyDalV5bYZtbYiw,740459
|
|
1667
|
+
wisent/parameters/lm_eval/group_task_evaluators.json,sha256=EzejSUN5nNl4cPd4HLMO9Mzmemdd7DKMYvvEv2I_B_M,47266
|
|
1668
|
+
wisent/parameters/lm_eval/group_tasks.json,sha256=mbbmDbMbggErQfT7We_eKreb11zbFvu7lB6MW007n8g,2483
|
|
1669
|
+
wisent/parameters/lm_eval/individual_tasks.json,sha256=dMfDQUlP48NTml0gwqMf8XVnMZco6VwdilrmnFCuDhs,6440
|
|
1670
|
+
wisent/parameters/lm_eval/no_readmes.json,sha256=T1PNoYwrqgwDVLtfmj7L5e0Sq02OEbqHPC8RFhICuUU,2
|
|
1671
|
+
wisent/parameters/lm_eval/not_lm_eval_tasks.json,sha256=zcdUnstl_OA52DO5Sf8ZAPkNcu9B-bfuyTXxPm8MEr8,1964
|
|
1672
|
+
wisent/parameters/lm_eval/read_tasks.json,sha256=MLOSbTPLXhMcPVw5QJosIIlUiRTFeQmbFZ16wFcMMKk,3401
|
|
1673
|
+
wisent/parameters/lm_eval/readme_files.json,sha256=1VtyYzREk50c1VjBz5p2JhUcydE0-Kv0wXmIR9C8X3U,3401
|
|
1674
|
+
wisent/parameters/lm_eval/track_progress_not_lm_eval_tasks.json,sha256=EdY5sllxx7l1YlzfOEKDqczJp8TNXM9ENHS2huqyWsQ,3792
|
|
1675
|
+
wisent/parameters/tasks/missing_task_families.json,sha256=h6g_2CYNQIW5SqRdcymDe4miWkGkVoD3yR0eOFZH4aw,65107
|
|
1676
|
+
wisent/parameters/tasks/remaining_tasks_to_implement.json,sha256=8_YDVQ439DdvC3daNW7v8y0Zth1eUxum6QzdnKDp6a8,3432
|
|
1677
|
+
wisent/parameters/tasks/risks.json,sha256=8U1LAY6Zrv68naYMzVlOrGgTBNL0l-9nC7moa-ljSHI,136
|
|
1678
|
+
wisent/parameters/tasks/skills.json,sha256=63KirwoeFTabMsQ3B7IdNyQLyl8857J0gJF02tTNPQo,215
|
|
1679
|
+
wisent/parameters/tasks/tasks.json,sha256=7HvSgYjO5dQfHfIwLB3D0h0lVQqpf2tKnU6U36j-r6Y,1140923
|
|
1680
|
+
wisent/scripts/run_quality_metrics_sweep.sh,sha256=NFqvIc1JB5fYYDXFBQeulQZV2cavbmsRQ2QWBU9aM10,10733
|
|
1681
|
+
wisent/tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1682
|
+
wisent/tests/test_all_cli_commands.py,sha256=7KqSZ8FsJEwXbgp0w_cfVtv9jgd-9P52fH3nx3jGF9k,23432
|
|
1683
|
+
wisent/tests/test_geometry_comprehensive.py,sha256=lUyBfZYudMMaX2-nV2QhtQcchbBaz8bLi6fqmMRQOig,13693
|
|
1684
|
+
wisent/tests/test_titan_geometry.py,sha256=lKWsFSCGEHFHD-CW_lz0vfATb6iFij0lrrLosBQJmkc,9710
|
|
1685
|
+
wisent/tests/visualize_geometry.py,sha256=mmv9bjA7MmjSZ1wS-J2h8G4g9YqtReOPmOE94giCibE,6119
|
|
1686
|
+
wisent/tests/examples/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1687
|
+
wisent/tests/examples/cli/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1688
|
+
wisent/tests/examples/cli/activations/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1689
|
+
wisent/tests/examples/cli/activations/test_get_activations.py,sha256=sIEpXyN4O3pPiIiZmbDAnKnRG2IubGl1L3M3D0agP_U,4009
|
|
1690
|
+
wisent/tests/examples/cli/classifier/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1691
|
+
wisent/tests/examples/cli/classifier/test_classifier_examples.py,sha256=kLSw-fxVl0auFFRW3J_KEEucndtwWefFkyRTn6z4G1o,4855
|
|
1692
|
+
wisent/tests/examples/cli/contrastive_pairs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1693
|
+
wisent/tests/examples/cli/contrastive_pairs/test_generate_pairs.py,sha256=oQv44hEYEGtk0DimgDbXJGrvyw_bsdJ-MBKvZUdXYAo,2860
|
|
1694
|
+
wisent/tests/examples/cli/evaluation/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1695
|
+
wisent/tests/examples/cli/evaluation/test_evaluation_examples.py,sha256=6YxRGtbPb08xFTjaBn1twz8_8OP7zMNM7ArxF8w4CyU,3551
|
|
1696
|
+
wisent/tests/examples/cli/generate/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1697
|
+
wisent/tests/examples/cli/generate/test_generate_with_classifier.py,sha256=FfUgHbeaeuexk61o0to9Cn6F47UjtVQyM1pUWRSm4mg,5096
|
|
1698
|
+
wisent/tests/examples/cli/generate/test_generate_with_steering.py,sha256=s-BmC-v8nkLFnCq42swo9gyFlT6Al3t_Dran3kswcoU,5183
|
|
1699
|
+
wisent/tests/examples/cli/generate/test_only_generate.py,sha256=WPuww-w9XtUWTLYMVgka2TFhkZWZ2EFZmLjQ5FZVjtg,3581
|
|
1700
|
+
wisent/tests/examples/cli/multi_steering/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1701
|
+
wisent/tests/examples/cli/multi_steering/test_multi_steer_from_trained_vectors.py,sha256=3MAD1aMGfP1s3Touczkl5mCbUmQuGuqnq8B7_y875dg,7537
|
|
1702
|
+
wisent/tests/examples/cli/multi_steering/test_multi_steer_with_different_parameters.py,sha256=vfUl6R7UsW5wADR5mleO4uGexKLrKocoUyhcIRVZCs4,6857
|
|
1703
|
+
wisent/tests/examples/cli/multi_steering/test_train_and_multi_steer.py,sha256=2nuvKhrIoCXxZfFoNWNs_V4Sifpuhiy2hAqmO8VJ2sg,6196
|
|
1704
|
+
wisent/tests/examples/cli/optimizer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1705
|
+
wisent/tests/examples/cli/optimizer/test_optimize_sample_size.py,sha256=3Matgc0vcuJFIL1pu7Tz_EVmmyOP77YeXDoo9viiMQM,3428
|
|
1706
|
+
wisent/tests/examples/cli/optimizer/test_optimizer_examples.py,sha256=l6L4eZDjmGWyUdkTcJGOK533LM1vlV7u0QcjmvjP4_o,1795
|
|
1707
|
+
wisent/tests/examples/cli/steering/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1708
|
+
wisent/tests/examples/cli/steering/test_create_steering_vectors.py,sha256=MJnUtRlsJqeteS33WU-_3fPW33PdLxT3tLBtRikmA34,4585
|
|
1709
|
+
wisent/tests/examples/cli/synthetic/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
1710
|
+
wisent/tests/examples/cli/synthetic/test_synthetic_pairs.py,sha256=d-OhvfMLMu_69T_RRNREGeiC-0aHJTooJZ6J4yR9ZKc,1320
|
|
1711
|
+
wisent/tests/nosense/__init__.py,sha256=sH3x4jRPzFM3YmQkdrwJoz-BdOQ1Bh6F95G5HWyIrv8,163
|
|
1712
|
+
wisent/tests/nosense/base_nosense.py,sha256=a18dBv1378nHly7OCIuk-bCcLnubss3XXDC1ex0zCK8,2633
|
|
1713
|
+
wisent/tests/nosense/math500_nosense.py,sha256=My0dHsr4OFOiTxb_VDKmGzpoMyzAtqXlHhA0oPfaG7s,2389
|
|
1714
|
+
wisent/tests/nosense/test_robustness.py,sha256=eeKji-_ls6tx7tuXqUO4BXxFRK-giJVihENAJVOvzSs,12546
|
|
1715
|
+
wisent-0.7.379.dist-info/licenses/LICENSE,sha256=wy0iaw8b2tyqZAfKHib3lP3PJ9o88FDCg92oUHh3sDQ,1073
|
|
1716
|
+
wisent-0.7.379.dist-info/METADATA,sha256=2e892Ck1O5N6u50diTk47KOO7SXkz09ssoJG1E02p2U,2125
|
|
1717
|
+
wisent-0.7.379.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
|
|
1718
|
+
wisent-0.7.379.dist-info/entry_points.txt,sha256=BM76j3xjtIcVZGk24iDf5w18s6SuqeOpaiAxfZhpnY8,49
|
|
1719
|
+
wisent-0.7.379.dist-info/top_level.txt,sha256=2Ts9Iyldnb3auIN2HBBaHPknRy7nSRDm2f6RGzYgr8A,7
|
|
1720
|
+
wisent-0.7.379.dist-info/RECORD,,
|