wisent 0.7.379__py3-none-any.whl → 0.7.701__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.
Files changed (725) hide show
  1. wisent/__init__.py +1 -1
  2. wisent/core/activations/__init__.py +22 -6
  3. wisent/core/activations/activations.py +21 -39
  4. wisent/core/activations/activations_collector.py +141 -373
  5. wisent/core/activations/classifier_inference_strategy.py +194 -0
  6. wisent/core/activations/core/atoms.py +8 -92
  7. wisent/core/activations/extraction_strategy.py +308 -0
  8. wisent/core/agent/diagnose/response_diagnostics.py +3 -3
  9. wisent/core/agent/diagnose.py +3 -3
  10. wisent/core/autonomous_agent.py +2 -2
  11. wisent/core/cli/agent/apply_steering.py +23 -27
  12. wisent/core/cli/agent/evaluate_response.py +18 -20
  13. wisent/core/cli/agent/train_classifier.py +18 -20
  14. wisent/core/cli/cluster_benchmarks.py +472 -0
  15. wisent/core/cli/create_steering_vector.py +13 -5
  16. wisent/core/cli/generate_vector_from_task.py +4 -0
  17. wisent/core/cli/get_activations.py +12 -36
  18. wisent/core/cli/method_optimizer.py +859 -0
  19. wisent/core/cli/optimize.py +44 -5
  20. wisent/core/cli/optimize_classification.py +5 -6
  21. wisent/core/cli/optimize_sample_size.py +8 -22
  22. wisent/core/cli/optimize_steering.py +429 -153
  23. wisent/core/cli/optimize_weights.py +65 -6
  24. wisent/core/cli/steering_method_trainer.py +5 -4
  25. wisent/core/cli/steering_search_space.py +20 -15
  26. wisent/core/cli/tasks.py +14 -43
  27. wisent/core/cli/train_unified_goodness.py +17 -18
  28. wisent/core/contrastive_pairs/diagnostics/control_vectors.py +1578 -173
  29. wisent/core/contrastive_pairs/diagnostics/linearity.py +63 -80
  30. wisent/core/contrastive_pairs/diagnostics/vector_quality.py +6 -5
  31. wisent/core/contrastive_pairs/huggingface_pairs/hf_extractor_manifest.py +5 -19
  32. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/__init__.py +11 -5
  33. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/apps.py +146 -32
  34. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue.py +2 -2
  35. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/humaneval.py +98 -57
  36. wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/code_x_glue.py +8 -8
  37. wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/freebase.py +1 -1
  38. wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_manifest.py +8 -5
  39. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/agieval_aqua_rat.py +129 -0
  40. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code_x_glue.py +11 -6
  41. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gsm8k.py +1 -1
  42. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mbpp.py +47 -6
  43. wisent/core/evaluators/benchmark_specific/apps_evaluator.py +133 -0
  44. wisent/core/evaluators/benchmark_specific/coding/metrics/evaluator.py +6 -1
  45. wisent/core/evaluators/benchmark_specific/conala_evaluator.py +31 -168
  46. wisent/core/evaluators/custom/examples/humanization_coherent.py +89 -35
  47. wisent/core/evaluators/oracles/truthfulqa_gen_evaluator.py +2 -20
  48. wisent/core/evaluators/personalization/coherence.py +46 -0
  49. wisent/core/hyperparameter_optimizer.py +13 -13
  50. wisent/core/lm_eval_harness_ground_truth.py +7 -11
  51. wisent/core/main.py +3 -0
  52. wisent/core/models/wisent_model.py +8 -7
  53. wisent/core/opti/methods/opti_weights.py +29 -2
  54. wisent/core/optuna/classifier/activation_generator.py +14 -12
  55. wisent/core/optuna/steering/steering_optimization.py +14 -9
  56. wisent/core/parser_arguments/cluster_benchmarks_parser.py +31 -0
  57. wisent/core/parser_arguments/generate_vector_from_task_parser.py +20 -0
  58. wisent/core/parser_arguments/main_parser.py +8 -0
  59. wisent/core/parser_arguments/optimize_steering_parser.py +117 -10
  60. wisent/core/parser_arguments/optimize_weights_parser.py +6 -0
  61. wisent/core/parser_arguments/tasks_parser.py +7 -19
  62. wisent/core/steering_methods/core/atoms.py +1 -2
  63. wisent/core/steering_methods/methods/caa.py +1 -1
  64. wisent/core/steering_methods/methods/hyperplane.py +74 -0
  65. wisent/core/steering_methods/methods/prism.py +1 -2
  66. wisent/core/steering_methods/methods/pulse.py +39 -8
  67. wisent/core/steering_methods/methods/titan.py +59 -14
  68. wisent/core/steering_methods/registry.py +52 -12
  69. wisent/core/steering_optimizer.py +15 -15
  70. wisent/core/trainers/steering_trainer.py +9 -18
  71. wisent/parameters/lm_eval/track_progress_not_lm_eval_tasks.json +19 -70
  72. wisent/scripts/run_quality_metrics_sweep.sh +22 -27
  73. wisent/tests/test_aggregation_geometry.py +236 -0
  74. wisent/tests/test_detector_accuracy.py +163 -0
  75. wisent/tests/test_geometry_exhaustive.py +1202 -0
  76. wisent/tests/visualize_geometry.py +255 -61
  77. {wisent-0.7.379.dist-info → wisent-0.7.701.dist-info}/METADATA +1 -1
  78. {wisent-0.7.379.dist-info → wisent-0.7.701.dist-info}/RECORD +82 -714
  79. wisent/core/activations/prompt_construction_strategy.py +0 -47
  80. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text.py +0 -15
  81. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_go.py +0 -64
  82. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_java.py +0 -65
  83. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_javascript.py +0 -65
  84. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_php.py +0 -65
  85. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_python.py +0 -65
  86. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue_code_to_text_ruby.py +0 -65
  87. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/freebase.py +0 -99
  88. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/instruct_humaneval.py +0 -180
  89. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/instructhumaneval.py +0 -129
  90. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mbpp.py +0 -142
  91. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/agieval.py +0 -155
  92. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code2text.py +0 -161
  93. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/codexglue.py +0 -107
  94. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/livemathbench.py +0 -155
  95. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/polymath.py +0 -155
  96. wisent/examples/scripts/results/benchmark_descriptions.json +0 -1244
  97. wisent/examples/scripts/results/benchmark_evaluation_methods.json +0 -66
  98. wisent/examples/scripts/results/benchmark_evaluator_mapping.json +0 -2781
  99. wisent/examples/scripts/results/benchmark_evaluator_mapping_updated.json +0 -30536
  100. wisent/examples/scripts/results/benchmark_evaluators_clean.json +0 -469
  101. wisent/examples/scripts/results/benchmark_methods_summary.json +0 -260
  102. wisent/examples/scripts/results/benchmark_pair_creation_methods.json +0 -66
  103. wisent/examples/scripts/results/benchmark_pair_totals.json +0 -269
  104. wisent/examples/scripts/results/benchmark_tags.json +0 -917
  105. wisent/examples/scripts/results/benchmark_test_summary_nov4.json +0 -71
  106. wisent/examples/scripts/results/coding_benchmarks_test_code_status.json +0 -150
  107. wisent/examples/scripts/results/failing_benchmarks.json +0 -946
  108. wisent/examples/scripts/results/failing_benchmarks_list.json +0 -41
  109. wisent/examples/scripts/results/failing_benchmarks_test_results.json +0 -945
  110. wisent/examples/scripts/results/missing_benchmark_tags.json +0 -341
  111. wisent/examples/scripts/results/test_20_newsgroups_evaluation.json +0 -30
  112. wisent/examples/scripts/results/test_20_newsgroups_pairs.json +0 -8
  113. wisent/examples/scripts/results/test_AraDICE_evaluation.json +0 -51
  114. wisent/examples/scripts/results/test_AraDICE_pairs.json +0 -14
  115. wisent/examples/scripts/results/test_AraDiCE_boolq_egy/test_AraDiCE_boolq_egy_evaluation.json +0 -30
  116. wisent/examples/scripts/results/test_AraDiCE_boolq_egy/test_AraDiCE_boolq_egy_pairs.json +0 -8
  117. wisent/examples/scripts/results/test_ArabCulture_evaluation.json +0 -51
  118. wisent/examples/scripts/results/test_ArabCulture_pairs.json +0 -14
  119. wisent/examples/scripts/results/test_Tag_evaluation.json +0 -30
  120. wisent/examples/scripts/results/test_Tag_pairs.json +0 -8
  121. wisent/examples/scripts/results/test_aclue_evaluation.json +0 -51
  122. wisent/examples/scripts/results/test_aclue_pairs.json +0 -14
  123. wisent/examples/scripts/results/test_acp_bench_evaluation.json +0 -51
  124. wisent/examples/scripts/results/test_acp_bench_hard_evaluation.json +0 -51
  125. wisent/examples/scripts/results/test_acp_bench_hard_pairs.json +0 -14
  126. wisent/examples/scripts/results/test_acp_bench_pairs.json +0 -14
  127. wisent/examples/scripts/results/test_advanced_ai_risk_evaluation.json +0 -51
  128. wisent/examples/scripts/results/test_advanced_ai_risk_pairs.json +0 -14
  129. wisent/examples/scripts/results/test_aexams_evaluation.json +0 -51
  130. wisent/examples/scripts/results/test_aexams_pairs.json +0 -14
  131. wisent/examples/scripts/results/test_afrimgsm_direct_amh_evaluation.json +0 -30
  132. wisent/examples/scripts/results/test_afrimgsm_direct_amh_pairs.json +0 -8
  133. wisent/examples/scripts/results/test_afrimmlu_direct_amh_evaluation.json +0 -30
  134. wisent/examples/scripts/results/test_afrimmlu_direct_amh_pairs.json +0 -8
  135. wisent/examples/scripts/results/test_afrixnli_en_direct_amh_evaluation.json +0 -30
  136. wisent/examples/scripts/results/test_afrixnli_en_direct_amh_pairs.json +0 -8
  137. wisent/examples/scripts/results/test_ag_news_evaluation.json +0 -30
  138. wisent/examples/scripts/results/test_ag_news_pairs.json +0 -8
  139. wisent/examples/scripts/results/test_agieval_evaluation.json +0 -51
  140. wisent/examples/scripts/results/test_agieval_pairs.json +0 -14
  141. wisent/examples/scripts/results/test_aime2024_evaluation.json +0 -30
  142. wisent/examples/scripts/results/test_aime2024_pairs.json +0 -8
  143. wisent/examples/scripts/results/test_aime2025_evaluation.json +0 -30
  144. wisent/examples/scripts/results/test_aime2025_pairs.json +0 -8
  145. wisent/examples/scripts/results/test_aime_evaluation.json +0 -30
  146. wisent/examples/scripts/results/test_aime_pairs.json +0 -8
  147. wisent/examples/scripts/results/test_anagrams1_evaluation.json +0 -30
  148. wisent/examples/scripts/results/test_anagrams1_pairs.json +0 -8
  149. wisent/examples/scripts/results/test_anagrams2_evaluation.json +0 -30
  150. wisent/examples/scripts/results/test_anagrams2_pairs.json +0 -8
  151. wisent/examples/scripts/results/test_anli_evaluation.json +0 -30
  152. wisent/examples/scripts/results/test_anli_pairs.json +0 -8
  153. wisent/examples/scripts/results/test_apps_evaluation.json +0 -30
  154. wisent/examples/scripts/results/test_apps_pairs.json +0 -8
  155. wisent/examples/scripts/results/test_arabic_exams_evaluation.json +0 -30
  156. wisent/examples/scripts/results/test_arabic_exams_pairs.json +0 -8
  157. wisent/examples/scripts/results/test_arabic_leaderboard_complete_evaluation.json +0 -51
  158. wisent/examples/scripts/results/test_arabic_leaderboard_complete_pairs.json +0 -14
  159. wisent/examples/scripts/results/test_arabic_leaderboard_light_evaluation.json +0 -51
  160. wisent/examples/scripts/results/test_arabic_leaderboard_light_pairs.json +0 -14
  161. wisent/examples/scripts/results/test_arabicmmlu_evaluation.json +0 -51
  162. wisent/examples/scripts/results/test_arabicmmlu_pairs.json +0 -14
  163. wisent/examples/scripts/results/test_aradice/test_aradice_evaluation.json +0 -51
  164. wisent/examples/scripts/results/test_aradice/test_aradice_pairs.json +0 -14
  165. wisent/examples/scripts/results/test_aradice3/test_aradice_evaluation.json +0 -51
  166. wisent/examples/scripts/results/test_aradice3/test_aradice_pairs.json +0 -14
  167. wisent/examples/scripts/results/test_arc_ar_evaluation.json +0 -30
  168. wisent/examples/scripts/results/test_arc_ar_pairs.json +0 -8
  169. wisent/examples/scripts/results/test_arc_challenge_evaluation.json +0 -30
  170. wisent/examples/scripts/results/test_arc_challenge_pairs.json +0 -8
  171. wisent/examples/scripts/results/test_arc_easy_evaluation.json +0 -30
  172. wisent/examples/scripts/results/test_arc_easy_pairs.json +0 -8
  173. wisent/examples/scripts/results/test_argument_topic_evaluation.json +0 -30
  174. wisent/examples/scripts/results/test_argument_topic_pairs.json +0 -8
  175. wisent/examples/scripts/results/test_arithmetic_evaluation.json +0 -51
  176. wisent/examples/scripts/results/test_arithmetic_pairs.json +0 -14
  177. wisent/examples/scripts/results/test_asdiv_evaluation.json +0 -30
  178. wisent/examples/scripts/results/test_asdiv_pairs.json +0 -8
  179. wisent/examples/scripts/results/test_assin_entailment_evaluation.json +0 -30
  180. wisent/examples/scripts/results/test_assin_entailment_pairs.json +0 -8
  181. wisent/examples/scripts/results/test_atis_evaluation.json +0 -30
  182. wisent/examples/scripts/results/test_atis_pairs.json +0 -8
  183. wisent/examples/scripts/results/test_babi_evaluation.json +0 -30
  184. wisent/examples/scripts/results/test_babi_pairs.json +0 -8
  185. wisent/examples/scripts/results/test_babilong_evaluation.json +0 -30
  186. wisent/examples/scripts/results/test_babilong_pairs.json +0 -8
  187. wisent/examples/scripts/results/test_bangla_mmlu_evaluation.json +0 -30
  188. wisent/examples/scripts/results/test_bangla_mmlu_pairs.json +0 -8
  189. wisent/examples/scripts/results/test_banking77_evaluation.json +0 -30
  190. wisent/examples/scripts/results/test_banking77_pairs.json +0 -8
  191. wisent/examples/scripts/results/test_basque/test_basque-glue_pairs.json +0 -14
  192. wisent/examples/scripts/results/test_basque-glue_evaluation.json +0 -51
  193. wisent/examples/scripts/results/test_basque-glue_pairs.json +0 -14
  194. wisent/examples/scripts/results/test_basque2/test_basque-glue_evaluation.json +0 -51
  195. wisent/examples/scripts/results/test_basque2/test_basque-glue_pairs.json +0 -14
  196. wisent/examples/scripts/results/test_basque_bench_evaluation.json +0 -51
  197. wisent/examples/scripts/results/test_basque_bench_pairs.json +0 -14
  198. wisent/examples/scripts/results/test_basque_glue/test_basque-glue_evaluation.json +0 -51
  199. wisent/examples/scripts/results/test_basque_glue/test_basque-glue_pairs.json +0 -14
  200. wisent/examples/scripts/results/test_basqueglue_evaluation.json +0 -51
  201. wisent/examples/scripts/results/test_basqueglue_pairs.json +0 -14
  202. wisent/examples/scripts/results/test_bbh_evaluation.json +0 -51
  203. wisent/examples/scripts/results/test_bbh_pairs.json +0 -14
  204. wisent/examples/scripts/results/test_bbq_evaluation.json +0 -30
  205. wisent/examples/scripts/results/test_bbq_pairs.json +0 -8
  206. wisent/examples/scripts/results/test_bec2016eu_evaluation.json +0 -51
  207. wisent/examples/scripts/results/test_bec2016eu_pairs.json +0 -14
  208. wisent/examples/scripts/results/test_belebele_evaluation.json +0 -51
  209. wisent/examples/scripts/results/test_belebele_pairs.json +0 -14
  210. wisent/examples/scripts/results/test_benchmarks_evaluation.json +0 -51
  211. wisent/examples/scripts/results/test_benchmarks_pairs.json +0 -14
  212. wisent/examples/scripts/results/test_bertaqa_evaluation.json +0 -51
  213. wisent/examples/scripts/results/test_bertaqa_pairs.json +0 -14
  214. wisent/examples/scripts/results/test_bhtc_v2_evaluation.json +0 -30
  215. wisent/examples/scripts/results/test_bhtc_v2_pairs.json +0 -8
  216. wisent/examples/scripts/results/test_bigbench_evaluation.json +0 -51
  217. wisent/examples/scripts/results/test_bigbench_pairs.json +0 -14
  218. wisent/examples/scripts/results/test_blimp_evaluation.json +0 -51
  219. wisent/examples/scripts/results/test_blimp_pairs.json +0 -14
  220. wisent/examples/scripts/results/test_boolq/test_boolq_evaluation.json +0 -30
  221. wisent/examples/scripts/results/test_boolq/test_boolq_pairs.json +0 -8
  222. wisent/examples/scripts/results/test_boolq-seq2seq_evaluation.json +0 -30
  223. wisent/examples/scripts/results/test_boolq-seq2seq_pairs.json +0 -8
  224. wisent/examples/scripts/results/test_boolq_evaluation.json +0 -30
  225. wisent/examples/scripts/results/test_boolq_pairs.json +0 -8
  226. wisent/examples/scripts/results/test_c4_evaluation.json +0 -30
  227. wisent/examples/scripts/results/test_c4_pairs.json +0 -8
  228. wisent/examples/scripts/results/test_cabreu_evaluation.json +0 -30
  229. wisent/examples/scripts/results/test_cabreu_pairs.json +0 -8
  230. wisent/examples/scripts/results/test_careqa_evaluation.json +0 -30
  231. wisent/examples/scripts/results/test_careqa_pairs.json +0 -8
  232. wisent/examples/scripts/results/test_catalan_bench_evaluation.json +0 -51
  233. wisent/examples/scripts/results/test_catalan_bench_pairs.json +0 -14
  234. wisent/examples/scripts/results/test_catalanqa_evaluation.json +0 -30
  235. wisent/examples/scripts/results/test_catalanqa_pairs.json +0 -8
  236. wisent/examples/scripts/results/test_catcola_evaluation.json +0 -30
  237. wisent/examples/scripts/results/test_catcola_pairs.json +0 -8
  238. wisent/examples/scripts/results/test_cb_evaluation.json +0 -30
  239. wisent/examples/scripts/results/test_cb_pairs.json +0 -8
  240. wisent/examples/scripts/results/test_ceval/test_ceval_evaluation.json +0 -51
  241. wisent/examples/scripts/results/test_ceval/test_ceval_pairs.json +0 -14
  242. wisent/examples/scripts/results/test_ceval_accountant/test_ceval-valid_accountant_evaluation.json +0 -30
  243. wisent/examples/scripts/results/test_ceval_accountant/test_ceval-valid_accountant_pairs.json +0 -8
  244. wisent/examples/scripts/results/test_ceval_evaluation.json +0 -51
  245. wisent/examples/scripts/results/test_ceval_pairs.json +0 -14
  246. wisent/examples/scripts/results/test_ceval_valid/test_ceval_valid_evaluation.json +0 -51
  247. wisent/examples/scripts/results/test_ceval_valid/test_ceval_valid_pairs.json +0 -14
  248. wisent/examples/scripts/results/test_chain_of_thought_evaluation.json +0 -51
  249. wisent/examples/scripts/results/test_chain_of_thought_pairs.json +0 -14
  250. wisent/examples/scripts/results/test_chartqa_evaluation.json +0 -30
  251. wisent/examples/scripts/results/test_chartqa_pairs.json +0 -8
  252. wisent/examples/scripts/results/test_claim_stance_topic_evaluation.json +0 -30
  253. wisent/examples/scripts/results/test_claim_stance_topic_pairs.json +0 -8
  254. wisent/examples/scripts/results/test_cmmlu_evaluation.json +0 -51
  255. wisent/examples/scripts/results/test_cmmlu_pairs.json +0 -14
  256. wisent/examples/scripts/results/test_cnn_dailymail_evaluation.json +0 -30
  257. wisent/examples/scripts/results/test_cnn_dailymail_pairs.json +0 -8
  258. wisent/examples/scripts/results/test_cocoteros_es_evaluation.json +0 -30
  259. wisent/examples/scripts/results/test_cocoteros_es_pairs.json +0 -8
  260. wisent/examples/scripts/results/test_codexglue_code_to_text_go_evaluation.json +0 -30
  261. wisent/examples/scripts/results/test_codexglue_code_to_text_go_pairs.json +0 -8
  262. wisent/examples/scripts/results/test_codexglue_code_to_text_java_evaluation.json +0 -30
  263. wisent/examples/scripts/results/test_codexglue_code_to_text_java_pairs.json +0 -8
  264. wisent/examples/scripts/results/test_codexglue_code_to_text_javascript_evaluation.json +0 -30
  265. wisent/examples/scripts/results/test_codexglue_code_to_text_javascript_pairs.json +0 -8
  266. wisent/examples/scripts/results/test_codexglue_code_to_text_php_evaluation.json +0 -30
  267. wisent/examples/scripts/results/test_codexglue_code_to_text_php_pairs.json +0 -8
  268. wisent/examples/scripts/results/test_codexglue_code_to_text_python_evaluation.json +0 -30
  269. wisent/examples/scripts/results/test_codexglue_code_to_text_python_pairs.json +0 -8
  270. wisent/examples/scripts/results/test_codexglue_code_to_text_ruby_evaluation.json +0 -30
  271. wisent/examples/scripts/results/test_codexglue_code_to_text_ruby_pairs.json +0 -8
  272. wisent/examples/scripts/results/test_coedit_gec_evaluation.json +0 -30
  273. wisent/examples/scripts/results/test_coedit_gec_pairs.json +0 -8
  274. wisent/examples/scripts/results/test_cola_evaluation.json +0 -30
  275. wisent/examples/scripts/results/test_cola_pairs.json +0 -8
  276. wisent/examples/scripts/results/test_commonsense_qa_evaluation.json +0 -30
  277. wisent/examples/scripts/results/test_commonsense_qa_pairs.json +0 -8
  278. wisent/examples/scripts/results/test_conala_evaluation.json +0 -30
  279. wisent/examples/scripts/results/test_conala_pairs.json +0 -8
  280. wisent/examples/scripts/results/test_concode_evaluation.json +0 -30
  281. wisent/examples/scripts/results/test_concode_pairs.json +0 -8
  282. wisent/examples/scripts/results/test_copa_evaluation.json +0 -30
  283. wisent/examples/scripts/results/test_copa_pairs.json +0 -8
  284. wisent/examples/scripts/results/test_copal_id_evaluation.json +0 -30
  285. wisent/examples/scripts/results/test_copal_id_pairs.json +0 -8
  286. wisent/examples/scripts/results/test_coqa_evaluation.json +0 -30
  287. wisent/examples/scripts/results/test_coqa_pairs.json +0 -8
  288. wisent/examples/scripts/results/test_coqcat_evaluation.json +0 -30
  289. wisent/examples/scripts/results/test_coqcat_pairs.json +0 -8
  290. wisent/examples/scripts/results/test_crows_pairs_evaluation.json +0 -51
  291. wisent/examples/scripts/results/test_crows_pairs_pairs.json +0 -14
  292. wisent/examples/scripts/results/test_csatqa_evaluation.json +0 -51
  293. wisent/examples/scripts/results/test_csatqa_pairs.json +0 -14
  294. wisent/examples/scripts/results/test_cycle_letters_evaluation.json +0 -30
  295. wisent/examples/scripts/results/test_cycle_letters_pairs.json +0 -8
  296. wisent/examples/scripts/results/test_darija_bench/test_darija_bench_evaluation.json +0 -51
  297. wisent/examples/scripts/results/test_darija_bench/test_darija_bench_pairs.json +0 -14
  298. wisent/examples/scripts/results/test_darija_bench_evaluation.json +0 -51
  299. wisent/examples/scripts/results/test_darija_bench_pairs.json +0 -14
  300. wisent/examples/scripts/results/test_darijahellaswag_evaluation.json +0 -30
  301. wisent/examples/scripts/results/test_darijahellaswag_pairs.json +0 -8
  302. wisent/examples/scripts/results/test_darijammlu_evaluation.json +0 -51
  303. wisent/examples/scripts/results/test_darijammlu_pairs.json +0 -14
  304. wisent/examples/scripts/results/test_dbpedia_14_evaluation.json +0 -30
  305. wisent/examples/scripts/results/test_dbpedia_14_pairs.json +0 -8
  306. wisent/examples/scripts/results/test_drop_evaluation.json +0 -30
  307. wisent/examples/scripts/results/test_drop_pairs.json +0 -8
  308. wisent/examples/scripts/results/test_ds1000_evaluation.json +0 -30
  309. wisent/examples/scripts/results/test_ds1000_pairs.json +0 -8
  310. wisent/examples/scripts/results/test_egyhellaswag_evaluation.json +0 -30
  311. wisent/examples/scripts/results/test_egyhellaswag_pairs.json +0 -8
  312. wisent/examples/scripts/results/test_egymmlu_evaluation.json +0 -51
  313. wisent/examples/scripts/results/test_egymmlu_pairs.json +0 -14
  314. wisent/examples/scripts/results/test_epec_koref_bin_evaluation.json +0 -30
  315. wisent/examples/scripts/results/test_epec_koref_bin_pairs.json +0 -8
  316. wisent/examples/scripts/results/test_eq_bench_evaluation.json +0 -30
  317. wisent/examples/scripts/results/test_eq_bench_pairs.json +0 -8
  318. wisent/examples/scripts/results/test_escola_evaluation.json +0 -30
  319. wisent/examples/scripts/results/test_escola_pairs.json +0 -8
  320. wisent/examples/scripts/results/test_ethics_cm_evaluation.json +0 -30
  321. wisent/examples/scripts/results/test_ethics_cm_pairs.json +0 -8
  322. wisent/examples/scripts/results/test_ethos_binary_evaluation.json +0 -30
  323. wisent/examples/scripts/results/test_ethos_binary_pairs.json +0 -8
  324. wisent/examples/scripts/results/test_eus_exams/test_eus_exams_evaluation.json +0 -51
  325. wisent/examples/scripts/results/test_eus_exams/test_eus_exams_pairs.json +0 -14
  326. wisent/examples/scripts/results/test_eus_exams_es_evaluation.json +0 -51
  327. wisent/examples/scripts/results/test_eus_exams_es_pairs.json +0 -14
  328. wisent/examples/scripts/results/test_eus_exams_evaluation.json +0 -51
  329. wisent/examples/scripts/results/test_eus_exams_pairs.json +0 -14
  330. wisent/examples/scripts/results/test_eus_proficiency_evaluation.json +0 -30
  331. wisent/examples/scripts/results/test_eus_proficiency_pairs.json +0 -8
  332. wisent/examples/scripts/results/test_eus_reading_evaluation.json +0 -30
  333. wisent/examples/scripts/results/test_eus_reading_pairs.json +0 -8
  334. wisent/examples/scripts/results/test_eus_trivia_evaluation.json +0 -30
  335. wisent/examples/scripts/results/test_eus_trivia_pairs.json +0 -8
  336. wisent/examples/scripts/results/test_evalita-mp_evaluation.json +0 -51
  337. wisent/examples/scripts/results/test_evalita-mp_pairs.json +0 -14
  338. wisent/examples/scripts/results/test_evalita-sp_sum_task_fp-small_p1_evaluation.json +0 -30
  339. wisent/examples/scripts/results/test_evalita-sp_sum_task_fp-small_p1_pairs.json +0 -8
  340. wisent/examples/scripts/results/test_evalita_LLM_evaluation.json +0 -51
  341. wisent/examples/scripts/results/test_evalita_LLM_pairs.json +0 -14
  342. wisent/examples/scripts/results/test_evalita_llm/test_evalita_llm_evaluation.json +0 -51
  343. wisent/examples/scripts/results/test_evalita_llm/test_evalita_llm_pairs.json +0 -14
  344. wisent/examples/scripts/results/test_evalita_mp/test_evalita-mp_te_prompt-1_evaluation.json +0 -30
  345. wisent/examples/scripts/results/test_evalita_mp/test_evalita-mp_te_prompt-1_pairs.json +0 -8
  346. wisent/examples/scripts/results/test_evalita_mp2/test_evalita_mp_evaluation.json +0 -51
  347. wisent/examples/scripts/results/test_evalita_mp2/test_evalita_mp_pairs.json +0 -14
  348. wisent/examples/scripts/results/test_evalita_sp2/test_evalita-sp_sum_task_fp-small_p1_evaluation.json +0 -30
  349. wisent/examples/scripts/results/test_evalita_sp2/test_evalita-sp_sum_task_fp-small_p1_pairs.json +0 -8
  350. wisent/examples/scripts/results/test_fda_evaluation.json +0 -30
  351. wisent/examples/scripts/results/test_fda_pairs.json +0 -8
  352. wisent/examples/scripts/results/test_financial_tweets_evaluation.json +0 -30
  353. wisent/examples/scripts/results/test_financial_tweets_pairs.json +0 -8
  354. wisent/examples/scripts/results/test_fld/test_fld_evaluation.json +0 -30
  355. wisent/examples/scripts/results/test_fld/test_fld_pairs.json +0 -8
  356. wisent/examples/scripts/results/test_fld_evaluation.json +0 -30
  357. wisent/examples/scripts/results/test_fld_fixed/test_fld_evaluation.json +0 -30
  358. wisent/examples/scripts/results/test_fld_fixed/test_fld_pairs.json +0 -8
  359. wisent/examples/scripts/results/test_fld_pairs.json +0 -8
  360. wisent/examples/scripts/results/test_flores_evaluation.json +0 -51
  361. wisent/examples/scripts/results/test_flores_pairs.json +0 -14
  362. wisent/examples/scripts/results/test_freebase_evaluation.json +0 -30
  363. wisent/examples/scripts/results/test_freebase_pairs.json +0 -8
  364. wisent/examples/scripts/results/test_french_bench_evaluation.json +0 -51
  365. wisent/examples/scripts/results/test_french_bench_pairs.json +0 -14
  366. wisent/examples/scripts/results/test_galcola_evaluation.json +0 -30
  367. wisent/examples/scripts/results/test_galcola_pairs.json +0 -8
  368. wisent/examples/scripts/results/test_galician_bench_evaluation.json +0 -51
  369. wisent/examples/scripts/results/test_galician_bench_pairs.json +0 -14
  370. wisent/examples/scripts/results/test_glianorex_evaluation.json +0 -30
  371. wisent/examples/scripts/results/test_glianorex_pairs.json +0 -8
  372. wisent/examples/scripts/results/test_global_mmlu_evaluation.json +0 -51
  373. wisent/examples/scripts/results/test_global_mmlu_pairs.json +0 -14
  374. wisent/examples/scripts/results/test_glue_evaluation.json +0 -51
  375. wisent/examples/scripts/results/test_glue_pairs.json +0 -14
  376. wisent/examples/scripts/results/test_gpqa_evaluation.json +0 -51
  377. wisent/examples/scripts/results/test_gpqa_pairs.json +0 -14
  378. wisent/examples/scripts/results/test_gpt3_translation_benchmarks_evaluation.json +0 -51
  379. wisent/examples/scripts/results/test_gpt3_translation_benchmarks_pairs.json +0 -14
  380. wisent/examples/scripts/results/test_groundcocoa_evaluation.json +0 -30
  381. wisent/examples/scripts/results/test_groundcocoa_pairs.json +0 -8
  382. wisent/examples/scripts/results/test_gsm8k_evaluation.json +0 -30
  383. wisent/examples/scripts/results/test_gsm8k_pairs.json +0 -8
  384. wisent/examples/scripts/results/test_haerae_evaluation.json +0 -51
  385. wisent/examples/scripts/results/test_haerae_pairs.json +0 -14
  386. wisent/examples/scripts/results/test_headqa_evaluation.json +0 -30
  387. wisent/examples/scripts/results/test_headqa_pairs.json +0 -8
  388. wisent/examples/scripts/results/test_hellaswag_evaluation.json +0 -30
  389. wisent/examples/scripts/results/test_hellaswag_pairs.json +0 -8
  390. wisent/examples/scripts/results/test_hendrycks_ethics_evaluation.json +0 -51
  391. wisent/examples/scripts/results/test_hendrycks_ethics_pairs.json +0 -14
  392. wisent/examples/scripts/results/test_hendrycks_math_evaluation.json +0 -51
  393. wisent/examples/scripts/results/test_hendrycks_math_pairs.json +0 -14
  394. wisent/examples/scripts/results/test_histoires_morales_evaluation.json +0 -30
  395. wisent/examples/scripts/results/test_histoires_morales_pairs.json +0 -8
  396. wisent/examples/scripts/results/test_hmmt_evaluation.json +0 -30
  397. wisent/examples/scripts/results/test_hmmt_feb_2025_evaluation.json +0 -30
  398. wisent/examples/scripts/results/test_hmmt_feb_2025_pairs.json +0 -8
  399. wisent/examples/scripts/results/test_hmmt_pairs.json +0 -8
  400. wisent/examples/scripts/results/test_hrm8k_evaluation.json +0 -51
  401. wisent/examples/scripts/results/test_hrm8k_pairs.json +0 -14
  402. wisent/examples/scripts/results/test_humaneval_evaluation.json +0 -30
  403. wisent/examples/scripts/results/test_humaneval_pairs.json +0 -8
  404. wisent/examples/scripts/results/test_humaneval_plus_evaluation.json +0 -30
  405. wisent/examples/scripts/results/test_humaneval_plus_pairs.json +0 -8
  406. wisent/examples/scripts/results/test_ifeval_evaluation.json +0 -30
  407. wisent/examples/scripts/results/test_ifeval_pairs.json +0 -8
  408. wisent/examples/scripts/results/test_instruct_humaneval/test_instruct_humaneval_evaluation.json +0 -30
  409. wisent/examples/scripts/results/test_instruct_humaneval/test_instruct_humaneval_pairs.json +0 -8
  410. wisent/examples/scripts/results/test_instruct_humaneval_evaluation.json +0 -30
  411. wisent/examples/scripts/results/test_instruct_humaneval_pairs.json +0 -8
  412. wisent/examples/scripts/results/test_inverse_scaling_evaluation.json +0 -51
  413. wisent/examples/scripts/results/test_inverse_scaling_hindsight_neglect_10shot_evaluation.json +0 -30
  414. wisent/examples/scripts/results/test_inverse_scaling_hindsight_neglect_10shot_pairs.json +0 -8
  415. wisent/examples/scripts/results/test_inverse_scaling_mc/test_inverse_scaling_mc_evaluation.json +0 -51
  416. wisent/examples/scripts/results/test_inverse_scaling_mc/test_inverse_scaling_mc_pairs.json +0 -14
  417. wisent/examples/scripts/results/test_inverse_scaling_pairs.json +0 -14
  418. wisent/examples/scripts/results/test_iwslt2017-ar-en_evaluation.json +0 -30
  419. wisent/examples/scripts/results/test_iwslt2017-ar-en_pairs.json +0 -8
  420. wisent/examples/scripts/results/test_iwslt2017-en-ar_evaluation.json +0 -30
  421. wisent/examples/scripts/results/test_iwslt2017-en-ar_pairs.json +0 -8
  422. wisent/examples/scripts/results/test_iwslt2017_ar_en/test_iwslt2017-ar-en_evaluation.json +0 -30
  423. wisent/examples/scripts/results/test_iwslt2017_ar_en/test_iwslt2017-ar-en_pairs.json +0 -8
  424. wisent/examples/scripts/results/test_iwslt2017_en_ar/test_iwslt2017-en-ar_evaluation.json +0 -30
  425. wisent/examples/scripts/results/test_iwslt2017_en_ar/test_iwslt2017-en-ar_pairs.json +0 -8
  426. wisent/examples/scripts/results/test_iwslt2017_group/test_iwslt2017_evaluation.json +0 -30
  427. wisent/examples/scripts/results/test_iwslt2017_group/test_iwslt2017_pairs.json +0 -8
  428. wisent/examples/scripts/results/test_japanese_leaderboard_evaluation.json +0 -51
  429. wisent/examples/scripts/results/test_japanese_leaderboard_pairs.json +0 -14
  430. wisent/examples/scripts/results/test_jsonschema_bench/test_jsonschema_bench_evaluation.json +0 -30
  431. wisent/examples/scripts/results/test_jsonschema_bench/test_jsonschema_bench_pairs.json +0 -8
  432. wisent/examples/scripts/results/test_jsonschema_bench_evaluation.json +0 -30
  433. wisent/examples/scripts/results/test_jsonschema_bench_final/test_jsonschema_bench_evaluation.json +0 -30
  434. wisent/examples/scripts/results/test_jsonschema_bench_final/test_jsonschema_bench_pairs.json +0 -8
  435. wisent/examples/scripts/results/test_jsonschema_bench_pairs.json +0 -8
  436. wisent/examples/scripts/results/test_kbl_evaluation.json +0 -51
  437. wisent/examples/scripts/results/test_kbl_fixed/test_kbl_evaluation.json +0 -51
  438. wisent/examples/scripts/results/test_kbl_fixed/test_kbl_pairs.json +0 -14
  439. wisent/examples/scripts/results/test_kbl_pairs.json +0 -14
  440. wisent/examples/scripts/results/test_kmmlu_evaluation.json +0 -51
  441. wisent/examples/scripts/results/test_kmmlu_pairs.json +0 -14
  442. wisent/examples/scripts/results/test_kobest_evaluation.json +0 -51
  443. wisent/examples/scripts/results/test_kobest_pairs.json +0 -14
  444. wisent/examples/scripts/results/test_kormedmcqa/test_kormedmcqa_evaluation.json +0 -30
  445. wisent/examples/scripts/results/test_kormedmcqa/test_kormedmcqa_pairs.json +0 -8
  446. wisent/examples/scripts/results/test_kormedmcqa_dentist/test_kormedmcqa_dentist_evaluation.json +0 -30
  447. wisent/examples/scripts/results/test_kormedmcqa_dentist/test_kormedmcqa_dentist_pairs.json +0 -8
  448. wisent/examples/scripts/results/test_kormedmcqa_evaluation.json +0 -30
  449. wisent/examples/scripts/results/test_kormedmcqa_pairs.json +0 -8
  450. wisent/examples/scripts/results/test_lambada_cloze_evaluation.json +0 -30
  451. wisent/examples/scripts/results/test_lambada_cloze_pairs.json +0 -8
  452. wisent/examples/scripts/results/test_lambada_evaluation.json +0 -30
  453. wisent/examples/scripts/results/test_lambada_final/test_lambada_openai_mt_stablelm_en_evaluation.json +0 -30
  454. wisent/examples/scripts/results/test_lambada_final/test_lambada_openai_mt_stablelm_en_pairs.json +0 -8
  455. wisent/examples/scripts/results/test_lambada_multilingual/test_lambada_multilingual_evaluation.json +0 -51
  456. wisent/examples/scripts/results/test_lambada_multilingual/test_lambada_multilingual_pairs.json +0 -14
  457. wisent/examples/scripts/results/test_lambada_multilingual_evaluation.json +0 -51
  458. wisent/examples/scripts/results/test_lambada_multilingual_pairs.json +0 -14
  459. wisent/examples/scripts/results/test_lambada_multilingual_stablelm_evaluation.json +0 -51
  460. wisent/examples/scripts/results/test_lambada_multilingual_stablelm_pairs.json +0 -14
  461. wisent/examples/scripts/results/test_lambada_openai_evaluation.json +0 -30
  462. wisent/examples/scripts/results/test_lambada_openai_pairs.json +0 -8
  463. wisent/examples/scripts/results/test_lambada_pairs.json +0 -8
  464. wisent/examples/scripts/results/test_lambada_stablelm_en_fixed/test_lambada_openai_mt_stablelm_en_evaluation.json +0 -30
  465. wisent/examples/scripts/results/test_lambada_stablelm_en_fixed/test_lambada_openai_mt_stablelm_en_pairs.json +0 -8
  466. wisent/examples/scripts/results/test_lambada_stablelm_fixed/test_lambada_openai_mt_stablelm_en_evaluation.json +0 -30
  467. wisent/examples/scripts/results/test_lambada_stablelm_fixed/test_lambada_openai_mt_stablelm_en_pairs.json +0 -8
  468. wisent/examples/scripts/results/test_lambada_standard_evaluation.json +0 -30
  469. wisent/examples/scripts/results/test_lambada_standard_pairs.json +0 -8
  470. wisent/examples/scripts/results/test_leaderboard_evaluation.json +0 -51
  471. wisent/examples/scripts/results/test_leaderboard_pairs.json +0 -14
  472. wisent/examples/scripts/results/test_libra/test_libra_evaluation.json +0 -51
  473. wisent/examples/scripts/results/test_libra/test_libra_pairs.json +0 -14
  474. wisent/examples/scripts/results/test_libra_evaluation.json +0 -51
  475. wisent/examples/scripts/results/test_libra_pairs.json +0 -14
  476. wisent/examples/scripts/results/test_lingoly_evaluation.json +0 -30
  477. wisent/examples/scripts/results/test_lingoly_pairs.json +0 -8
  478. wisent/examples/scripts/results/test_livecodebench_evaluation.json +0 -30
  479. wisent/examples/scripts/results/test_livecodebench_pairs.json +0 -8
  480. wisent/examples/scripts/results/test_livemathbench_cnmo_en_evaluation.json +0 -30
  481. wisent/examples/scripts/results/test_livemathbench_cnmo_en_pairs.json +0 -8
  482. wisent/examples/scripts/results/test_livemathbench_cnmo_zh_evaluation.json +0 -30
  483. wisent/examples/scripts/results/test_livemathbench_cnmo_zh_pairs.json +0 -8
  484. wisent/examples/scripts/results/test_llama_evaluation.json +0 -30
  485. wisent/examples/scripts/results/test_llama_pairs.json +0 -8
  486. wisent/examples/scripts/results/test_logiqa2_evaluation.json +0 -30
  487. wisent/examples/scripts/results/test_logiqa2_pairs.json +0 -8
  488. wisent/examples/scripts/results/test_logiqa_evaluation.json +0 -30
  489. wisent/examples/scripts/results/test_logiqa_pairs.json +0 -8
  490. wisent/examples/scripts/results/test_m_mmlu_evaluation.json +0 -51
  491. wisent/examples/scripts/results/test_m_mmlu_pairs.json +0 -14
  492. wisent/examples/scripts/results/test_mastermind/test_mastermind_evaluation.json +0 -51
  493. wisent/examples/scripts/results/test_mastermind/test_mastermind_pairs.json +0 -14
  494. wisent/examples/scripts/results/test_mastermind_24_easy/test_mastermind_24_easy_evaluation.json +0 -30
  495. wisent/examples/scripts/results/test_mastermind_24_easy/test_mastermind_24_easy_pairs.json +0 -8
  496. wisent/examples/scripts/results/test_mastermind_evaluation.json +0 -51
  497. wisent/examples/scripts/results/test_mastermind_pairs.json +0 -14
  498. wisent/examples/scripts/results/test_math500_evaluation.json +0 -30
  499. wisent/examples/scripts/results/test_math500_pairs.json +0 -8
  500. wisent/examples/scripts/results/test_math_evaluation.json +0 -30
  501. wisent/examples/scripts/results/test_math_pairs.json +0 -8
  502. wisent/examples/scripts/results/test_mathqa_evaluation.json +0 -30
  503. wisent/examples/scripts/results/test_mathqa_pairs.json +0 -8
  504. wisent/examples/scripts/results/test_mbpp_evaluation.json +0 -30
  505. wisent/examples/scripts/results/test_mbpp_pairs.json +0 -8
  506. wisent/examples/scripts/results/test_mbpp_plus_evaluation.json +0 -30
  507. wisent/examples/scripts/results/test_mbpp_plus_pairs.json +0 -8
  508. wisent/examples/scripts/results/test_mc_taco_evaluation.json +0 -30
  509. wisent/examples/scripts/results/test_mc_taco_pairs.json +0 -8
  510. wisent/examples/scripts/results/test_med_concepts_qa/test_med_concepts_qa_evaluation.json +0 -51
  511. wisent/examples/scripts/results/test_med_concepts_qa/test_med_concepts_qa_pairs.json +0 -14
  512. wisent/examples/scripts/results/test_med_concepts_qa_atc_easy/test_med_concepts_qa_atc_easy_evaluation.json +0 -30
  513. wisent/examples/scripts/results/test_med_concepts_qa_atc_easy/test_med_concepts_qa_atc_easy_pairs.json +0 -8
  514. wisent/examples/scripts/results/test_med_concepts_qa_evaluation.json +0 -51
  515. wisent/examples/scripts/results/test_med_concepts_qa_pairs.json +0 -14
  516. wisent/examples/scripts/results/test_meddialog_evaluation.json +0 -30
  517. wisent/examples/scripts/results/test_meddialog_pairs.json +0 -8
  518. wisent/examples/scripts/results/test_meddialog_raw_perplexity/test_meddialog_raw_perplexity_evaluation.json +0 -30
  519. wisent/examples/scripts/results/test_meddialog_raw_perplexity/test_meddialog_raw_perplexity_pairs.json +0 -8
  520. wisent/examples/scripts/results/test_mediqa_qa2019_evaluation.json +0 -30
  521. wisent/examples/scripts/results/test_mediqa_qa2019_pairs.json +0 -8
  522. wisent/examples/scripts/results/test_medmcqa_evaluation.json +0 -30
  523. wisent/examples/scripts/results/test_medmcqa_pairs.json +0 -8
  524. wisent/examples/scripts/results/test_medqa_evaluation.json +0 -30
  525. wisent/examples/scripts/results/test_medqa_pairs.json +0 -8
  526. wisent/examples/scripts/results/test_medtext_evaluation.json +0 -30
  527. wisent/examples/scripts/results/test_medtext_pairs.json +0 -8
  528. wisent/examples/scripts/results/test_mela_evaluation.json +0 -51
  529. wisent/examples/scripts/results/test_mela_pairs.json +0 -14
  530. wisent/examples/scripts/results/test_meqsum_evaluation.json +0 -30
  531. wisent/examples/scripts/results/test_meqsum_pairs.json +0 -8
  532. wisent/examples/scripts/results/test_mercury_evaluation.json +0 -30
  533. wisent/examples/scripts/results/test_mercury_pairs.json +0 -8
  534. wisent/examples/scripts/results/test_metabench_evaluation.json +0 -51
  535. wisent/examples/scripts/results/test_metabench_pairs.json +0 -14
  536. wisent/examples/scripts/results/test_mgsm_evaluation.json +0 -51
  537. wisent/examples/scripts/results/test_mgsm_pairs.json +0 -14
  538. wisent/examples/scripts/results/test_mimic_repsum_evaluation.json +0 -30
  539. wisent/examples/scripts/results/test_mimic_repsum_pairs.json +0 -8
  540. wisent/examples/scripts/results/test_minerva_math_evaluation.json +0 -51
  541. wisent/examples/scripts/results/test_minerva_math_pairs.json +0 -14
  542. wisent/examples/scripts/results/test_mlqa_evaluation.json +0 -51
  543. wisent/examples/scripts/results/test_mlqa_pairs.json +0 -14
  544. wisent/examples/scripts/results/test_mmlu-pro-plus_evaluation.json +0 -51
  545. wisent/examples/scripts/results/test_mmlu-pro-plus_pairs.json +0 -14
  546. wisent/examples/scripts/results/test_mmlu_evaluation.json +0 -51
  547. wisent/examples/scripts/results/test_mmlu_pairs.json +0 -14
  548. wisent/examples/scripts/results/test_mmlu_pro_evaluation.json +0 -51
  549. wisent/examples/scripts/results/test_mmlu_pro_pairs.json +0 -14
  550. wisent/examples/scripts/results/test_mmlu_prox_evaluation.json +0 -51
  551. wisent/examples/scripts/results/test_mmlu_prox_pairs.json +0 -14
  552. wisent/examples/scripts/results/test_mmlusr_evaluation.json +0 -30
  553. wisent/examples/scripts/results/test_mmlusr_pairs.json +0 -8
  554. wisent/examples/scripts/results/test_mmmu_evaluation.json +0 -51
  555. wisent/examples/scripts/results/test_mmmu_pairs.json +0 -14
  556. wisent/examples/scripts/results/test_mnli_evaluation.json +0 -30
  557. wisent/examples/scripts/results/test_mnli_pairs.json +0 -8
  558. wisent/examples/scripts/results/test_model_written_evals_evaluation.json +0 -51
  559. wisent/examples/scripts/results/test_model_written_evals_pairs.json +0 -14
  560. wisent/examples/scripts/results/test_moral_stories_evaluation.json +0 -30
  561. wisent/examples/scripts/results/test_moral_stories_pairs.json +0 -8
  562. wisent/examples/scripts/results/test_mts_dialog_evaluation.json +0 -30
  563. wisent/examples/scripts/results/test_mts_dialog_pairs.json +0 -8
  564. wisent/examples/scripts/results/test_multiblimp_evaluation.json +0 -51
  565. wisent/examples/scripts/results/test_multiblimp_pairs.json +0 -14
  566. wisent/examples/scripts/results/test_multimedqa_evaluation.json +0 -51
  567. wisent/examples/scripts/results/test_multimedqa_pairs.json +0 -14
  568. wisent/examples/scripts/results/test_multipl_e_evaluation.json +0 -30
  569. wisent/examples/scripts/results/test_multipl_e_pairs.json +0 -8
  570. wisent/examples/scripts/results/test_mutual_evaluation.json +0 -30
  571. wisent/examples/scripts/results/test_mutual_pairs.json +0 -8
  572. wisent/examples/scripts/results/test_non_greedy_robustness_agieval_aqua_rat_evaluation.json +0 -30
  573. wisent/examples/scripts/results/test_non_greedy_robustness_agieval_aqua_rat_pairs.json +0 -8
  574. wisent/examples/scripts/results/test_noreval_evaluation.json +0 -51
  575. wisent/examples/scripts/results/test_noreval_pairs.json +0 -14
  576. wisent/examples/scripts/results/test_noticia_evaluation.json +0 -30
  577. wisent/examples/scripts/results/test_noticia_pairs.json +0 -8
  578. wisent/examples/scripts/results/test_nq_open_evaluation.json +0 -30
  579. wisent/examples/scripts/results/test_nq_open_pairs.json +0 -8
  580. wisent/examples/scripts/results/test_olaph_evaluation.json +0 -30
  581. wisent/examples/scripts/results/test_olaph_pairs.json +0 -8
  582. wisent/examples/scripts/results/test_openbookqa_evaluation.json +0 -30
  583. wisent/examples/scripts/results/test_openbookqa_pairs.json +0 -8
  584. wisent/examples/scripts/results/test_openllm_evaluation.json +0 -51
  585. wisent/examples/scripts/results/test_openllm_pairs.json +0 -14
  586. wisent/examples/scripts/results/test_option_order_robustness_agieval_aqua_rat_evaluation.json +0 -30
  587. wisent/examples/scripts/results/test_option_order_robustness_agieval_aqua_rat_pairs.json +0 -8
  588. wisent/examples/scripts/results/test_paloma_evaluation.json +0 -51
  589. wisent/examples/scripts/results/test_paloma_pairs.json +0 -14
  590. wisent/examples/scripts/results/test_passkey/test_passkey_evaluation.json +0 -30
  591. wisent/examples/scripts/results/test_passkey/test_passkey_pairs.json +0 -8
  592. wisent/examples/scripts/results/test_paws-x_evaluation.json +0 -51
  593. wisent/examples/scripts/results/test_paws-x_pairs.json +0 -14
  594. wisent/examples/scripts/results/test_paws_en/test_paws_en_evaluation.json +0 -30
  595. wisent/examples/scripts/results/test_paws_en/test_paws_en_pairs.json +0 -8
  596. wisent/examples/scripts/results/test_penn_treebank_evaluation.json +0 -30
  597. wisent/examples/scripts/results/test_penn_treebank_pairs.json +0 -8
  598. wisent/examples/scripts/results/test_pile_10k/test_pile_10k_evaluation.json +0 -30
  599. wisent/examples/scripts/results/test_pile_10k/test_pile_10k_pairs.json +0 -8
  600. wisent/examples/scripts/results/test_piqa_evaluation.json +0 -30
  601. wisent/examples/scripts/results/test_piqa_pairs.json +0 -8
  602. wisent/examples/scripts/results/test_polemo2_evaluation.json +0 -30
  603. wisent/examples/scripts/results/test_polemo2_pairs.json +0 -8
  604. wisent/examples/scripts/results/test_polymath_en_high_evaluation.json +0 -30
  605. wisent/examples/scripts/results/test_polymath_en_high_pairs.json +0 -8
  606. wisent/examples/scripts/results/test_polymath_en_medium_evaluation.json +0 -30
  607. wisent/examples/scripts/results/test_polymath_en_medium_pairs.json +0 -8
  608. wisent/examples/scripts/results/test_polymath_zh_high_evaluation.json +0 -30
  609. wisent/examples/scripts/results/test_polymath_zh_high_pairs.json +0 -8
  610. wisent/examples/scripts/results/test_polymath_zh_medium_evaluation.json +0 -30
  611. wisent/examples/scripts/results/test_polymath_zh_medium_pairs.json +0 -8
  612. wisent/examples/scripts/results/test_portuguese_bench_evaluation.json +0 -51
  613. wisent/examples/scripts/results/test_portuguese_bench_pairs.json +0 -14
  614. wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat/test_prompt_robustness_agieval_aqua_rat_evaluation.json +0 -30
  615. wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat/test_prompt_robustness_agieval_aqua_rat_pairs.json +0 -8
  616. wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat_evaluation.json +0 -30
  617. wisent/examples/scripts/results/test_prompt_robustness_agieval_aqua_rat_pairs.json +0 -8
  618. wisent/examples/scripts/results/test_prost_evaluation.json +0 -30
  619. wisent/examples/scripts/results/test_prost_pairs.json +0 -8
  620. wisent/examples/scripts/results/test_ptb_evaluation.json +0 -30
  621. wisent/examples/scripts/results/test_ptb_pairs.json +0 -8
  622. wisent/examples/scripts/results/test_pubmedqa_evaluation.json +0 -30
  623. wisent/examples/scripts/results/test_pubmedqa_pairs.json +0 -8
  624. wisent/examples/scripts/results/test_pythia_evaluation.json +0 -51
  625. wisent/examples/scripts/results/test_pythia_pairs.json +0 -14
  626. wisent/examples/scripts/results/test_qa4mre_evaluation.json +0 -30
  627. wisent/examples/scripts/results/test_qa4mre_pairs.json +0 -8
  628. wisent/examples/scripts/results/test_qasper_evaluation.json +0 -30
  629. wisent/examples/scripts/results/test_qasper_pairs.json +0 -8
  630. wisent/examples/scripts/results/test_race_evaluation.json +0 -30
  631. wisent/examples/scripts/results/test_race_pairs.json +0 -8
  632. wisent/examples/scripts/results/test_realtoxicityprompts_evaluation.json +0 -30
  633. wisent/examples/scripts/results/test_realtoxicityprompts_pairs.json +0 -8
  634. wisent/examples/scripts/results/test_recode_evaluation.json +0 -30
  635. wisent/examples/scripts/results/test_recode_pairs.json +0 -8
  636. wisent/examples/scripts/results/test_record_evaluation.json +0 -30
  637. wisent/examples/scripts/results/test_record_pairs.json +0 -8
  638. wisent/examples/scripts/results/test_ruler_evaluation.json +0 -51
  639. wisent/examples/scripts/results/test_ruler_pairs.json +0 -14
  640. wisent/examples/scripts/results/test_sciq_evaluation.json +0 -30
  641. wisent/examples/scripts/results/test_sciq_pairs.json +0 -8
  642. wisent/examples/scripts/results/test_score_evaluation.json +0 -51
  643. wisent/examples/scripts/results/test_score_pairs.json +0 -14
  644. wisent/examples/scripts/results/test_self_consistency_evaluation.json +0 -30
  645. wisent/examples/scripts/results/test_self_consistency_pairs.json +0 -8
  646. wisent/examples/scripts/results/test_siqa/test_siqa_evaluation.json +0 -30
  647. wisent/examples/scripts/results/test_siqa/test_siqa_pairs.json +0 -8
  648. wisent/examples/scripts/results/test_siqa_evaluation.json +0 -30
  649. wisent/examples/scripts/results/test_siqa_pairs.json +0 -8
  650. wisent/examples/scripts/results/test_spanish_bench_evaluation.json +0 -51
  651. wisent/examples/scripts/results/test_spanish_bench_pairs.json +0 -14
  652. wisent/examples/scripts/results/test_squad2_evaluation.json +0 -30
  653. wisent/examples/scripts/results/test_squad2_pairs.json +0 -8
  654. wisent/examples/scripts/results/test_squadv2_evaluation.json +0 -30
  655. wisent/examples/scripts/results/test_squadv2_pairs.json +0 -8
  656. wisent/examples/scripts/results/test_super-glue-lm-eval-v1-seq2seq_evaluation.json +0 -30
  657. wisent/examples/scripts/results/test_super-glue-lm-eval-v1-seq2seq_pairs.json +0 -8
  658. wisent/examples/scripts/results/test_super-glue-lm-eval-v1_evaluation.json +0 -51
  659. wisent/examples/scripts/results/test_super-glue-lm-eval-v1_pairs.json +0 -14
  660. wisent/examples/scripts/results/test_swag_evaluation.json +0 -30
  661. wisent/examples/scripts/results/test_swag_pairs.json +0 -8
  662. wisent/examples/scripts/results/test_tinyBenchmarks_evaluation.json +0 -51
  663. wisent/examples/scripts/results/test_tinyBenchmarks_pairs.json +0 -14
  664. wisent/examples/scripts/results/test_tmmluplus_evaluation.json +0 -51
  665. wisent/examples/scripts/results/test_tmmluplus_pairs.json +0 -14
  666. wisent/examples/scripts/results/test_translation_evaluation.json +0 -51
  667. wisent/examples/scripts/results/test_translation_pairs.json +0 -14
  668. wisent/examples/scripts/results/test_triviaqa_evaluation.json +0 -30
  669. wisent/examples/scripts/results/test_triviaqa_pairs.json +0 -8
  670. wisent/examples/scripts/results/test_truthfulqa-multi_evaluation.json +0 -51
  671. wisent/examples/scripts/results/test_truthfulqa-multi_pairs.json +0 -14
  672. wisent/examples/scripts/results/test_truthfulqa_evaluation.json +0 -30
  673. wisent/examples/scripts/results/test_truthfulqa_mc1_evaluation.json +0 -30
  674. wisent/examples/scripts/results/test_truthfulqa_mc1_pairs.json +0 -8
  675. wisent/examples/scripts/results/test_truthfulqa_mc2_evaluation.json +0 -30
  676. wisent/examples/scripts/results/test_truthfulqa_mc2_pairs.json +0 -8
  677. wisent/examples/scripts/results/test_truthfulqa_pairs.json +0 -8
  678. wisent/examples/scripts/results/test_turkishmmlu_evaluation.json +0 -51
  679. wisent/examples/scripts/results/test_turkishmmlu_pairs.json +0 -14
  680. wisent/examples/scripts/results/test_unfair_tos_evaluation.json +0 -30
  681. wisent/examples/scripts/results/test_unfair_tos_pairs.json +0 -8
  682. wisent/examples/scripts/results/test_unscramble_evaluation.json +0 -51
  683. wisent/examples/scripts/results/test_unscramble_pairs.json +0 -14
  684. wisent/examples/scripts/results/test_webqs_evaluation.json +0 -30
  685. wisent/examples/scripts/results/test_webqs_pairs.json +0 -8
  686. wisent/examples/scripts/results/test_wikitext103_evaluation.json +0 -30
  687. wisent/examples/scripts/results/test_wikitext103_pairs.json +0 -8
  688. wisent/examples/scripts/results/test_wikitext_evaluation.json +0 -30
  689. wisent/examples/scripts/results/test_wikitext_pairs.json +0 -8
  690. wisent/examples/scripts/results/test_winogender_evaluation.json +0 -51
  691. wisent/examples/scripts/results/test_winogender_pairs.json +0 -14
  692. wisent/examples/scripts/results/test_winogrande_evaluation.json +0 -30
  693. wisent/examples/scripts/results/test_winogrande_pairs.json +0 -8
  694. wisent/examples/scripts/results/test_wmdp_evaluation.json +0 -30
  695. wisent/examples/scripts/results/test_wmdp_pairs.json +0 -8
  696. wisent/examples/scripts/results/test_wmt-ro-en-t5-prompt_evaluation.json +0 -30
  697. wisent/examples/scripts/results/test_wmt-ro-en-t5-prompt_pairs.json +0 -8
  698. wisent/examples/scripts/results/test_wmt14_en_fr_evaluation.json +0 -30
  699. wisent/examples/scripts/results/test_wmt14_en_fr_pairs.json +0 -8
  700. wisent/examples/scripts/results/test_wmt16_en_de_evaluation.json +0 -30
  701. wisent/examples/scripts/results/test_wmt16_en_de_pairs.json +0 -8
  702. wisent/examples/scripts/results/test_wmt16_ro_en_evaluation.json +0 -30
  703. wisent/examples/scripts/results/test_wmt16_ro_en_pairs.json +0 -8
  704. wisent/examples/scripts/results/test_wsc273_evaluation.json +0 -30
  705. wisent/examples/scripts/results/test_wsc273_pairs.json +0 -8
  706. wisent/examples/scripts/results/test_xcopa_evaluation.json +0 -51
  707. wisent/examples/scripts/results/test_xcopa_pairs.json +0 -14
  708. wisent/examples/scripts/results/test_xnli_eu_evaluation.json +0 -30
  709. wisent/examples/scripts/results/test_xnli_eu_pairs.json +0 -8
  710. wisent/examples/scripts/results/test_xnli_evaluation.json +0 -51
  711. wisent/examples/scripts/results/test_xnli_pairs.json +0 -14
  712. wisent/examples/scripts/results/test_xquad_evaluation.json +0 -51
  713. wisent/examples/scripts/results/test_xquad_pairs.json +0 -14
  714. wisent/examples/scripts/results/test_xstorycloze_evaluation.json +0 -51
  715. wisent/examples/scripts/results/test_xstorycloze_pairs.json +0 -14
  716. wisent/examples/scripts/results/test_xsum_evaluation.json +0 -30
  717. wisent/examples/scripts/results/test_xsum_pairs.json +0 -8
  718. wisent/examples/scripts/results/test_xwinograd_evaluation.json +0 -51
  719. wisent/examples/scripts/results/test_xwinograd_pairs.json +0 -14
  720. wisent/examples/scripts/results/test_yahoo_answers_topics_evaluation.json +0 -30
  721. wisent/examples/scripts/results/test_yahoo_answers_topics_pairs.json +0 -8
  722. {wisent-0.7.379.dist-info → wisent-0.7.701.dist-info}/WHEEL +0 -0
  723. {wisent-0.7.379.dist-info → wisent-0.7.701.dist-info}/entry_points.txt +0 -0
  724. {wisent-0.7.379.dist-info → wisent-0.7.701.dist-info}/licenses/LICENSE +0 -0
  725. {wisent-0.7.379.dist-info → wisent-0.7.701.dist-info}/top_level.txt +0 -0
{wisent-0.7.379.dist-info → wisent-0.7.701.dist-info}/RECORD RENAMED
@@ -1,7 +1,7 @@
1
- wisent/__init__.py,sha256=jXMUyFIIcSsWnNECcJrbJnDD33bboO3fY9GF9q8bKTU,1229
1
+ wisent/__init__.py,sha256=AG1M4u2B5EcGHVjK-FMR_6hNR3P9q1OzxP3E11LExQY,1229
2
2
  wisent/cli.py,sha256=XKzGIGstr38EowHYpr821c6YuV9Eaw3I1I3NvLztTO0,3960
3
3
  wisent/core/__init__.py,sha256=x1MX4vKpKP3c2FuIHcFly-UkoZwGVnRPbzcFaxr_Jdo,1340
4
- wisent/core/autonomous_agent.py,sha256=n0j2YSmhAfLt2WBaKM4o33SNQiiLYjvIRqfnT1nbO4M,48681
4
+ wisent/core/autonomous_agent.py,sha256=rtJ5XlWqPys7c1PBFkUBHGg9Ox6_LDjeoiNPEph6NA4,48667
5
5
  wisent/core/benchmark_extractors.py,sha256=BtGoL23n9EzSGoywymSh8uIhlr8T_LU8y-hp7et1vIk,13616
6
6
  wisent/core/benchmark_registry.py,sha256=KQWMDEKNLBRvzd91HyH8SlhwIKXwpJtTWoSPcIuTn5Y,4627
7
7
  wisent/core/bigcode_extractors.py,sha256=xKrJLdglcpykCv9TR72PVwu4cq3Fp5W6XuMYrdDHsVc,696
@@ -12,10 +12,10 @@ wisent/core/config_manager.py,sha256=RfB9PT3K1Ns5EsQiSGQ1DSxdlG-cVMdaaUuqQMdCNOY
12
12
  wisent/core/detection_handling.py,sha256=L14US_y4UWaSabeAEzWn6tPM79qQFgwpO-BXvBBO7fA,11183
13
13
  wisent/core/diversity_processors.py,sha256=wm4rFK8qrvYLcDpaJVHe-r6QtDYDRGty-Eyd0nXzcdM,6883
14
14
  wisent/core/download_full_benchmarks.py,sha256=6JVmof4xKC--Lkj8r6v70KGRxPVTqu7YC3nVsb5m42g,60722
15
- wisent/core/hyperparameter_optimizer.py,sha256=OHkk_uP91QSeysAyzCFLzGqJdbfH3cIUxJT1kWA-PpY,22987
15
+ wisent/core/hyperparameter_optimizer.py,sha256=_aMhzjUlXGGMQeSsX8gWXWf-zwyTopA-mbbOzPEnbSA,22761
16
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
17
+ wisent/core/lm_eval_harness_ground_truth.py,sha256=46WGnsJLoQ3ALXnXJ0PkTo-N1U2wGDGKyU1kwg-1Ap0,66514
18
+ wisent/core/main.py,sha256=dovHLbcRV8zEwKz7nRmryJntBS8QODCzI78Kw9UpBps,4353
19
19
  wisent/core/managed_cached_benchmarks.py,sha256=_q26-13t_GLFvkMA3M8tge0Sv6fFJmH_AJqCVl-Dt3w,23499
20
20
  wisent/core/mixed_benchmark_sampler.py,sha256=J1R7Q7LOXE28P2Bydx_NrgblJv35rqdkCPkvZiAxp_c,13715
21
21
  wisent/core/model_persistence.py,sha256=Pr1A3E0MxqwICKsFXmjKVq9KgG8Ip5UbtG-gujGFShc,9870
@@ -25,7 +25,7 @@ wisent/core/representation.py,sha256=hBl_N9qbr5Gsa7GCQ0nMWRm82RqYEfhd9cyf0PPH5LY
25
25
  wisent/core/save_results.py,sha256=NJ7elQzbh3Q_ZefdTV8L-9QyutteWSBWfgTw78OBm1k,13725
26
26
  wisent/core/steering.py,sha256=oiDIvHVm-n0eCS_Er9kv-Bu0H2JOqns87QhY5Oz-T2A,22790
27
27
  wisent/core/steering_method.py,sha256=Y09jIImhZnRuHkbpeAfJT1hDovXi0bacqt3X3TVSo68,411
28
- wisent/core/steering_optimizer.py,sha256=OdgOCLnEYUnjAiP19f_I80167K8W08CrScVhYyDlvuw,65587
28
+ wisent/core/steering_optimizer.py,sha256=SVMI6xHtMb6xAZYkXggziqjEZ9SSkWCL0PiBj4WMlsc,65353
29
29
  wisent/core/task_interface.py,sha256=4xm6nrduFdXSZpJccahfZWtpI4rZvCQcsS7GCJWpbMU,4906
30
30
  wisent/core/task_selector.py,sha256=JLMQ8E1HAvs4H4UqJRIvvWaktJsSAo3uz8oY135qO9g,7774
31
31
  wisent/core/time_estimator.py,sha256=Frwwbg8DRaeFlxBNewhLPhveysfOZQDR7oMhcJO6kt0,5963
@@ -33,12 +33,13 @@ wisent/core/timing_calibration.py,sha256=zIu1lgIEX_lBAmVfHakOHSEOvUAtFI5y4i3sk5e
33
33
  wisent/core/universal_subspace.py,sha256=D35PgZE42IwzjdyrTjmPcYou9YKJkbVZPfwCHh2jF1g,30728
34
34
  wisent/core/user_model_config.py,sha256=FGzHYo1kD7qDPhpffvDkigV-z7aiLxHAtqiMEPf_sIM,6729
35
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
36
+ wisent/core/activations/__init__.py,sha256=KjKMOPbPMmi1iMQMH6ALkcwa8pZLK8dpbz0cHGVjl_M,1352
37
+ wisent/core/activations/activations.py,sha256=htpxdaF18uxknd6ETs_3iBKwWqh3bI3x6JwB-jkfk6E,3175
38
+ wisent/core/activations/activations_collector.py,sha256=WPLX0-wVV-HbYTyaXo3okqiiWD72mQ6k5JDC85hpgM4,10214
39
+ wisent/core/activations/classifier_inference_strategy.py,sha256=oA8VEbuWKLQPZUvjbXC1_f6jAvLVCII3DHLHvXKJBHo,6910
40
+ wisent/core/activations/extraction_strategy.py,sha256=N2m8OWobcOGFDsngan1PA2w6S1zhozwG9bnT_vu8vjo,12432
40
41
  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/activations/core/atoms.py,sha256=wfPhgSt7NP_4xkyYZA6T62KTM2DUkq1Jj-FJeRd9Kko,5383
42
43
  wisent/core/adapters/__init__.py,sha256=UZkssKYL4-Jl_F29x29MQr6L4899lpK-eFvFa_qGa8Y,712
43
44
  wisent/core/adapters/audio.py,sha256=hqEXJ5NSfGstnQLmYkaBw_4TZ9Ft2qJaBMOKtObkb4Q,21712
44
45
  wisent/core/adapters/base.py,sha256=ir75AwPA39sjpzGIs3aPFL2poZ34CVqVmygO-KZHbpA,12824
@@ -49,14 +50,14 @@ wisent/core/adapters/video.py,sha256=M1p2rHFTNkN2LjteYXp3NL3L6v1E238XMR8Agv7O5wI
49
50
  wisent/core/agent/__init__.py,sha256=cAzpae7HWdNrpxCqANWUA2TKkqpm_hoSj8JAdbfmLwM,81
50
51
  wisent/core/agent/budget.py,sha256=zBhMTvdVGLqM6vIhHzm5BO2dAp_ANuJvFPmZ8_9YQdM,24301
51
52
  wisent/core/agent/device_benchmarks.py,sha256=2G_SuiM87gAS0TPZaMzS03yTOTHlbwJzpWNxYH60IJY,26206
52
- wisent/core/agent/diagnose.py,sha256=8WOFmFI5nrXvo81egt3nPiYgufK1mGqQKaUDfFwWbZc,9324
53
+ wisent/core/agent/diagnose.py,sha256=HpPnPDt0UE9yXLJeuWZdL5W38FoXRKHI0S2st05l4C4,9298
53
54
  wisent/core/agent/steer.py,sha256=4v3kkWTo4qerEJSa-D2CBUQE0ESQRN90Rd3ovL48uQo,9928
54
55
  wisent/core/agent/timeout.py,sha256=WvYC9n9huf0b5daXX1vt2G4b3OEsU7irPd15nj_Twbc,4328
55
56
  wisent/core/agent/diagnose/__init__.py,sha256=cAzpae7HWdNrpxCqANWUA2TKkqpm_hoSj8JAdbfmLwM,81
56
57
  wisent/core/agent/diagnose/agent_classifier_decision.py,sha256=WrrIW-ig4V_OoVQUHhieENOhEMjy0t72Z-ElgxWQS5s,28978
57
58
  wisent/core/agent/diagnose/classifier_marketplace.py,sha256=m02pwQkC4ROui6BhNhGJArdgBREI2xYuaRinm09VxM0,22113
58
59
  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/response_diagnostics.py,sha256=pEqxZVXtRJoadrJeFanWRzne_hIfeRHk4cEuPo_j0t0,10629
60
61
  wisent/core/agent/diagnose/select_classifiers.py,sha256=XmmG3eAYkMNoPGE55-0gQTIGtDeqc_5sg93Mx03r__k,18353
61
62
  wisent/core/agent/diagnose/synthetic_classifier_option.py,sha256=ZWsyZM7iUoFPlDtyWDjxi1qOeoVo9bUqFtl6eZ5pqzM,32703
62
63
  wisent/core/agent/diagnose/tasks/__init__.py,sha256=IGnEhyeCWNagVjRpIDoX4Msb1rxfeWLrsPa3EvhiYeI,698
@@ -78,7 +79,8 @@ wisent/core/classifiers/models/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVx
78
79
  wisent/core/classifiers/pipeline_steps/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
79
80
  wisent/core/cli/__init__.py,sha256=0paqfodiH1pMUIzqLbnc9HB-reukpjfwzmgmfUrOMX4,1929
80
81
  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/cluster_benchmarks.py,sha256=1k0rOO8jgEop4hWvkJqPNqNmSnmYeQEYBT2QXHaMezo,20163
83
+ wisent/core/cli/create_steering_vector.py,sha256=M11P5WkmEiuflR7qNRPA_pPkkZAQcbe7fsg5P4ddirs,16640
82
84
  wisent/core/cli/diagnose_pairs.py,sha256=dopxJUgxien33yx6zAhwRVfb_4uvIem7Zpj2lawC1dQ,6775
83
85
  wisent/core/cli/diagnose_vectors.py,sha256=jWLyk990j7jHtP1XVga6hVaR6LOzF6pzjwhkq--GZfs,17088
84
86
  wisent/core/cli/estimate_unified_goodness_time.py,sha256=KnT7qWBkKwn7M1iaM62NRQpKdF_3pmoTQq4w7WYbRRU,16510
@@ -89,28 +91,29 @@ wisent/core/cli/generate_pairs.py,sha256=KT88DaRMv5UcQwkLY9hf8uF3vuq9i3pWSnAL6Ne
89
91
  wisent/core/cli/generate_pairs_from_task.py,sha256=YppLsiyDJhv4U7aoMBggLshLD5MsRQGirrLrg6oBLLc,4406
90
92
  wisent/core/cli/generate_responses.py,sha256=nPEwQ8_EVjRbr0lhsCl8sacQzvt5XgDUOONSwRyI--U,5238
91
93
  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/generate_vector_from_task.py,sha256=O619eALpDrTWMq4Milih_IC3eDViXz46EP6D1A6uduE,10121
95
+ wisent/core/cli/get_activations.py,sha256=tXEvtIIWv-98vyched-3MJORdy5HUfelWmE4M8Fx8gI,6612
94
96
  wisent/core/cli/inference_config.py,sha256=iIZa0_ciBbM0Fqat0DJNd-mlWd0ANuYoeZ1v1V4jduc,3148
95
97
  wisent/core/cli/inference_config_cli.py,sha256=CKvD3TpFYDSNtm9HdRFv0uMGInaR6iSHP8aLWYPhrfs,1786
98
+ wisent/core/cli/method_optimizer.py,sha256=J17rwW9ZHxXZwWTkkgW7HfOQ_rRynvS9Wd2CqWyuT0I,32791
96
99
  wisent/core/cli/modify_weights.py,sha256=JhZwUl91lWsZwdiie60FQCxAfcwewtICBXf8XOLlBdw,28562
97
100
  wisent/core/cli/multi_steer.py,sha256=GRs0557YabcEwch7P75uWJadHf4Ssa8YE5nTMt8WTvs,4564
98
101
  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
102
+ wisent/core/cli/optimize.py,sha256=4ZpLawTrvU57SUEEhxYiR-Xy5c3mZrE9gSRIxgSRWSg,28090
103
+ wisent/core/cli/optimize_classification.py,sha256=deg3psSPvERDptRIfv4x88adKL0RM3gCBJGfb0QdWnI,23099
104
+ wisent/core/cli/optimize_sample_size.py,sha256=EQpIA-lYkh74x3JKRLlmec0RyVElnC-KwYkqUx1H6kM,16204
105
+ wisent/core/cli/optimize_steering.py,sha256=UyBVszZPCjqPu45VdROiksyof2ycauZ0YQ8kRP8EgCk,173041
106
+ wisent/core/cli/optimize_weights.py,sha256=2uybF7B7LB28-mhTzb088OeWOeX90H_9CHj1MEWYK9k,54285
107
+ wisent/core/cli/steering_method_trainer.py,sha256=UKkur2Ice8zXJwho6TG9G9aO_Lk99G2bcscOUycoVA8,22887
108
+ wisent/core/cli/steering_search_space.py,sha256=4iGf7vTHPEE4JW1mQdq_P1vLYWyGlUekoyjMMJgx0Hw,23113
109
+ wisent/core/cli/tasks.py,sha256=jyqTOfAQjmuUwS0guRzLjxbmUgLZ0PPbXtgTbAoEORs,43381
110
+ wisent/core/cli/train_unified_goodness.py,sha256=rX6rRVYYM66WDvIexiLe-FALBHGuns2kRY94zEsZTmE,29086
108
111
  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
112
+ wisent/core/cli/agent/apply_steering.py,sha256=HDeX6nhhO87z1BzNdZeaAH-1B-K4twvRNIkoQEKlQRE,6729
113
+ wisent/core/cli/agent/evaluate_response.py,sha256=Xk-nKSeGxTqMX7IF6tjYdTUkc_EITVIUg8XEYqZBSnk,4522
111
114
  wisent/core/cli/agent/generate_synthetic_pairs.py,sha256=QnMuJvunEB3Icmx9TrFeMklh1fjpTMLeu8mkV_dx46w,4723
112
115
  wisent/core/cli/agent/main.py,sha256=aZ7SIuzTmYkFTaygjTf76Fgho9CiRMIOooes5Vd7ueI,5457
113
- wisent/core/cli/agent/train_classifier.py,sha256=E8z3FXk2ZiWNtBZ8Q1o1SZ3XTag2ZajY7t9nXUsTQ0c,6934
116
+ wisent/core/cli/agent/train_classifier.py,sha256=zWqZkVb4wZQnJfUXa8Jvk9C-Xbeo8VOU-kCtH4TVs_c,6545
114
117
  wisent/core/contrastive_pairs/__init__.py,sha256=AbaAf-t_nyVVy_vLjp8WAlMDmNun3KNp_GMWAK25r9g,429
115
118
  wisent/core/contrastive_pairs/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
116
119
  wisent/core/contrastive_pairs/core/atoms.py,sha256=_zghw6c8iisW_SqBIUCoAnzhc5q7t5EgZ4zzTPxeLwQ,1129
@@ -122,17 +125,17 @@ wisent/core/contrastive_pairs/core/set.py,sha256=sVJ2zvlVaFpMPgr4dYe-FMhT6sPS0KO
122
125
  wisent/core/contrastive_pairs/diagnostics/__init__.py,sha256=X17XJthKEO4gUxUqBR01ysPwf7jx4LZoPKu3QZPK8Xk,2366
123
126
  wisent/core/contrastive_pairs/diagnostics/activations.py,sha256=TUNxU-HV3oeQxFbZhKo-OISM4mzR-Bib0naHmbqWzk8,1736
124
127
  wisent/core/contrastive_pairs/diagnostics/base.py,sha256=uBi8PdTd6BRyy0lmGjAZLTZdgiiWwPNtsmKkBFCmlD0,2658
125
- wisent/core/contrastive_pairs/diagnostics/control_vectors.py,sha256=yLFtaXI9Fw2vjKpzDepCd6H_91g8G96V3xhToyHhzd0,63131
128
+ wisent/core/contrastive_pairs/diagnostics/control_vectors.py,sha256=C3QCDHDgrL7LJxW40qMXGtEptmLuZ-oL1etclNmOIrU,120062
126
129
  wisent/core/contrastive_pairs/diagnostics/coverage.py,sha256=MpT6_IdLqtMpav6mOCiNuemBVFvxWzkUbj7j3ZNx-48,2761
127
130
  wisent/core/contrastive_pairs/diagnostics/divergence.py,sha256=Io3AcGluJogz4qENWu0ivQyFR_5bLN49BzCTI7DIVa4,3430
128
131
  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
132
+ wisent/core/contrastive_pairs/diagnostics/linearity.py,sha256=ElTjTYub0txZamCf1IPunYShTCv1VdfHdQqGhPse4NA,10842
133
+ wisent/core/contrastive_pairs/diagnostics/vector_quality.py,sha256=EmgH05y5j9EbdiNxCc3FuhIvCM1DoUsGCgQVMkYSjYk,22311
131
134
  wisent/core/contrastive_pairs/huggingface_pairs/__init__.py,sha256=PYur_GOluHmwaYUS6rxsoas5-AmrG7OhVGoPteuJB-c,76
132
135
  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
136
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_extractor_manifest.py,sha256=GwgoH0zXKQ9J29aT_taZzPrLYTJBpTc82fqjHKEOsbc,26345
134
137
  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
138
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/__init__.py,sha256=HtFG8o1n69SkHgbkjqPXXqsPMPlRUzFu6UWqQ76c268,2522
136
139
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentbench.py,sha256=tdJznfmBU0PxtHwzgXVIJy5qRCGPQ898i_-Mu2FFoF8,7936
137
140
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentharm.py,sha256=hWEPf8mKIgZDZuisG0YMtAr04c33lfXjkRT8-wr2zfo,10504
138
141
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentic_search.py,sha256=fKSzaCY5Y2mRjRYr2OpP8RDQ6GN6LU3DHepq3rlNV30,17563
@@ -141,7 +144,7 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime.py,sha25
141
144
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime2024.py,sha256=iWtVKnwr1dS4k-zDU9nS4qsN3hnoFNjKb4JOIHsRuqM,2234
142
145
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime2025.py,sha256=l2NC4sA7yCpsEx-NtyQmj7x19KfXHcprsYhxVA8a6OQ,2132
143
146
  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
147
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/apps.py,sha256=SEMU4cKx5BQUMrKTn7MISece_c2NJDv-PehSW0cmI6k,10281
145
148
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/arena_hard.py,sha256=T9qLzUpFLHZz1OBwGJ4_-7c6_3ZWswDebbvYle3sINw,6824
146
149
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/atis.py,sha256=UJfzAett3soThQGvnkFctninkNPMK5YFj16SVm0dsdw,3102
147
150
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/babilong.py,sha256=sqXHWyyRE1LMZB5IZtSdldxP7PP6OR890bZ68Vn4wt8,3050
@@ -156,14 +159,7 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/chinese_simpl
156
159
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/cluewsc.py,sha256=yhi43rtNT_zjfZ2yg2BuxHI0kSDSETKc-tPk9eeFqgc,6300
157
160
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/cnn_dailymail.py,sha256=NSWDMh2ScUsvC_WVAgIWfGkkY64F_lPbbj9U4FokD7s,3200
158
161
  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
162
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue.py,sha256=voaKHYSQV91mkY9pqHehNFqO-nnAKxuq5T0r4JRVsm4,3352
167
163
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coding_benchmarks.py,sha256=QXJBlUj_rJIwXqn1uNouyT76hBgwOHMYJabaLIfE9ng,25867
168
164
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coedit_gec.py,sha256=a-0kJQFcAmoLNFE-dV-G07th9qXPQE0IuAXPcZaVEm8,2605
169
165
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/conala.py,sha256=auULgkRhG77vy2rAHJlpRhAsdTLrIZuq_XKK9PZF9tU,4075
@@ -184,7 +180,6 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flames.py,sha
184
180
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flan_held_in.py,sha256=Xzx2V6j8ihoYwC4SYG226SXXEH5j9G2vJFOMT0Y7pR8,3650
185
181
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flores.py,sha256=MwEv7rWS1_e9PMdmC88SM7yNkh189ZJgbhQXKMvmjT0,23149
186
182
  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
183
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/get_negative_example_livecodebench.py,sha256=lrhwIQtQupD-xM9ypOamMcPuohkYHIb5uoDvmvgF8JA,5129
189
184
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/get_positive_example_livecodebench.py,sha256=k_PSaWtokvvBfysz3goFh-ZneXGSAKDaLQ22kkMl5oQ,4622
190
185
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/gpt3_translation_benchmarks.py,sha256=aOsJ-l67Noc1qnq_orp7-iw8wSNLX-VCa9C_gXX7MAc,3773
@@ -194,10 +189,8 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/harmbench.py,
194
189
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/healthbench.py,sha256=_APNLxYsMgu3rJigXJwWdB83qL-X3pBMRJF0fzAPWNo,6831
195
190
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hle.py,sha256=awu6NcomLtATCwps67BQZJ0bZNDTx9EtYgoi9dshVlU,3176
196
191
  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
192
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/humaneval.py,sha256=OWJnf6AsY4vcvZixlWZi2ui2pb_kqiPBzAIWxrHH0Dg,4797
198
193
  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
194
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/iwslt2017_ar_en.py,sha256=OakHs-5Qli9tt9gXXsQ913u1aJqbNm4bP7hQ3kbAol0,3826
202
195
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/iwslt2017_en_ar.py,sha256=GszqSy52EgJ1L885KnhvA5zs_TEBbb6_NJP__aTXWgI,3826
203
196
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/jailbreakbench.py,sha256=DI9kig2nHMgLZ-Ow4EpwYTQlpNinTW1ul15ms_r4SJA,9642
@@ -212,7 +205,6 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/longform_writ
212
205
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/m_mmlu.py,sha256=RDZmJAZfJdpxXwrbhDPFBOClNDV2gMf9TaBWNhVF3Gw,3724
213
206
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math.py,sha256=1HUMRBIN9b26aWOcyM70tEBbLfzIiZ9uw3OmKZFBpEA,5980
214
207
  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
208
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/meddialog.py,sha256=n-Mh0iWhd-v6avJVAIKe3juHn_D5CrqsXCgERQNK8uA,2616
217
209
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medical_abstracts.py,sha256=TP1jcbcVRaDuwwSI_CusIwlchI2QWfMhsCa3rHLH_RY,3576
218
210
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medium_priority_benchmarks.py,sha256=-NHzHBPM8I0zVUkdCO4gsN0RyZ3vKJQHyP9RUNfKOso,30169
@@ -276,7 +268,7 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt_ro_en_t5_
276
268
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/xsum.py,sha256=zzIp0MeqI7XccHlNEaqL0w3OR3Ukt6lqMYVZQmksAMs,2907
277
269
  wisent/core/contrastive_pairs/lm_eval_pairs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
278
270
  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
271
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_manifest.py,sha256=GUvxCEOKl06gzjl_UTTOsCWUfTXs5MiB1IsQQukY_e8,30200
280
272
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_registry.py,sha256=Wq5_AOWRUO2m387YaLmzWPWfMmwT96FmaAz0N0_JV6A,5255
281
273
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_pairs_generation.py,sha256=x5dWdceSnMetwmxliUFmErAi0DHXknSBUQ3SSXBznh4,2090
282
274
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/__init__.py,sha256=yxkFjHDEpD33FH1-shObI-jsperBSADjTqZxQrzC3QA,16627
@@ -329,7 +321,7 @@ wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/careqa.py,sha25
329
321
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/catalan_bench.py,sha256=TPUzqtL1du59Dt9EoHaKriyoFromWMAoAswjy52lOeM,3167
330
322
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/ceval_valid.py,sha256=UNjahQmaNAA4Tbk1VRrwjwskbh62xZjNTcCh75Agu_c,4229
331
323
  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
324
+ wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/code_x_glue.py,sha256=Fx-tjjXV01dOJUWp_J0ylLaBsDUYSkpQKz3InC0AKr8,706
333
325
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/copal_id.py,sha256=he7n0aGC0XjMR4qa2gm0kuovyz0_x1JLEYJNPfXDmuA,372
334
326
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/crows_pairs.py,sha256=7CTdKKc494cmFzBjofRyp2_dlNwiyqBPiuwlkDyXAjY,2136
335
327
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/csatqa.py,sha256=Eo22i1KWxQX-PNZ3nGPm3UQSKmP8oxrY3V8zchNwOrc,571
@@ -340,7 +332,7 @@ wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/eus.py,sha256=i
340
332
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/evalita_mp.py,sha256=_LdM__wODPNdavTHGgLqF3dA_vLHC-ewmydHm8Bs7q8,6644
341
333
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/fld.py,sha256=FLzUNm80tvs-htbAfm4mNq-HSiu8oEYufTzstf9Nk4o,210
342
334
  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
335
+ wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/freebase.py,sha256=RtCWXKQwDa70bd3lSnmPAcIWrIBu_QxR2uKHnR9bvMc,226
344
336
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/french_bench.py,sha256=Rqlk55ooPSoKOp1vQCT-jPsDqinDAC6ot5YkCIy5XwA,1554
345
337
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/galician_bench.py,sha256=csNGQJO-ML0ViEuvWGDaa-V22tdOylnx0cPsSbbTdXU,3073
346
338
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/glianorex.py,sha256=diTVUzsCmpkEl-xfnMCmrUChExkEqRPtQpZ2qHmN-Q0,372
@@ -444,7 +436,7 @@ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrixnli.py,sha25
444
436
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrobench_cot.py,sha256=XXg_BHyIHOt_NezgVJ5K8aplTR_rTiX1Eg154L3sTI0,3319
445
437
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrobench_mc.py,sha256=1aRM-_bxDOqWWc5WCfaes_WIbgfs-zurYQkP_PxR660,4114
446
438
  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
439
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/agieval_aqua_rat.py,sha256=oUJD14n4sty8M2b8WXtH84XrC1GJD3r_lqMrZXHL0sQ,4637
448
440
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ai2_arc.py,sha256=0Rrn2efrqGl949i30Y83R450uoc_y8o4H31qj46dr3g,3878
449
441
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anagrams1.py,sha256=VMEhdDEjXyQscQGSXz98H42LZ75XTuS-QcrARzWx-QQ,3141
450
442
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anagrams2.py,sha256=gnfUcTl26pUmoYPGDg5SUBasrmCK1ggB_kIQYsijNeU,3141
@@ -500,9 +492,7 @@ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/click.py,sha256=x
500
492
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cmmlu.py,sha256=jvoB7X6gi0-tByNHlIMuIaLvQdLMW3WrsNATVuNWi8w,6661
501
493
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cnn.py,sha256=o3CkInmKzwZZDNFtNtRotFQaTnQrCsUjn19A1I8EWhE,5838
502
494
  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
495
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code_x_glue.py,sha256=D14EUwniYnfUP7Z5mU5zRfkPAd1bCtYwEMd0tAavAE0,4511
506
496
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coedit.py,sha256=jHan7PwlX3wyQjsW49qzAw-P7DlJPYjZQ-tEoNiU-l8,5875
507
497
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cola.py,sha256=dRnZPR63cbrzVyiklwJX05cMB2fUTujZhNpQC6IVG6c,2996
508
498
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense.py,sha256=QLNyRHimX8aVUiNuT0K--I5-vYo-FNFTnvTfGjNNGJM,4296
@@ -560,7 +550,7 @@ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpqa.py,sha256=On
560
550
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpt3.py,sha256=jLlJFYXCiO20kWI22OO64rxWjp1W9kjjD13ygx9DNE0,4354
561
551
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/groundcocoa.py,sha256=x8PbYJ_TvyeEKj33JD-CfyuR23jzCVNUuEX-zvj-8t8,7246
562
552
  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
553
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gsm8k.py,sha256=zGkwC5Tevv7B8RBA-3D9EOykU_XLoJ7BNfiZrtX0vlE,4598
564
554
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/haerae.py,sha256=4AUa0l3HgFqj-Lw5UvmRr6gV2b29ZgLyngLK7MS3nTE,5947
565
555
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/headqa.py,sha256=6XAzuoqSDahyjX5BjE23zTUlv3KN5_PHlu0fjRIsrs8,3875
566
556
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hellaswag.py,sha256=EFwXe4TAQsJtSq5U3I7yQK6_9GG_K2quG5PrQc-xJNo,4412
@@ -595,7 +585,6 @@ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/law.py,sha256=96H
595
585
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/leaderboard.py,sha256=nPRQZ9996-Jc1yrnElMyebbyZdyUA9ssnG2t7nhUqbM,8075
596
586
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/libra.py,sha256=le4wRHUc2Dpq0gTzS_lQKCZ8zBvlYMVFq0OMkzkdrck,6107
597
587
  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
588
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/llama3.py,sha256=RiQLmKvG24XtH7ryrquAexP2Bqc1z16WV3gmsurrcAM,5947
600
589
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lm_syneval.py,sha256=3N2Bt-aAaxugVmQGwt0MdQZ1qKzm8i_qWuqeQJ28zZ8,5981
601
590
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logieval.py,sha256=b8mWP5I0RUsmLSDE0UyYg9VC87C_qGQWn8Uq78G_dAE,2932
@@ -605,7 +594,7 @@ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbench.py,sha2
605
594
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbenchv2.py,sha256=5f89b7cg4NHa_1vsIbFFz-osh3srP_4KMfbCzWsHc9E,5992
606
595
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mastermind.py,sha256=iyXStISnO5I9KBkgy9sh0nC9eAxmubZ4Iu0J8hIbga8,8019
607
596
  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
597
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mbpp.py,sha256=qkCmDbZcE5gOYMMeJrvoxjtZxq9wa7bHiEGpi1Sp1EE,6035
609
598
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mc-taco.py,sha256=rSvVod1RH3CWYzZi1hH4MJOnYvnFYFwhOSbwQhckSJk,3955
610
599
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/med_concepts_qa.py,sha256=QbCTKrOXGmd6oP7Pt-yRa8idy5ZcIROuXLsjsSeouRM,8785
611
600
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meddialog.py,sha256=ld97Bu1mJneVcQ1RlcEyJvnRwxLCJUfYReHKSlMwv9Q,7082
@@ -668,7 +657,6 @@ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile.py,sha256=cf
668
657
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile_10k.py,sha256=K5diNndk9obpncxS2h2pJ3WeFpceZTlCO7d1QXeKCYA,4845
669
658
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/piqa.py,sha256=iXN0CxFzLDvLt8pWyuV9MlOUXIiIIkN7obg79eNPtnE,3884
670
659
  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
660
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench.py,sha256=Jg3foz6stbTDqr9Fi2B1YUd2ktH_PRvfi3-VfCazs98,6105
673
661
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench_gen.py,sha256=hGLtogrXR_yib9riVQZAkqpwCsMbhb7F8K-AR0uKFgU,4191
674
662
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench_mc.py,sha256=u02jIQXUOuwbjfCrDtd1rIEvPq4e9CPvvsa7De8DSYk,4133
@@ -777,7 +765,8 @@ wisent/core/evaluators/steering_evaluators.py,sha256=3b2voLO9BzRzNmEjhiw2GzOuHLy
777
765
  wisent/core/evaluators/synthetic_evaluator.py,sha256=odKzYAHp1pIIvZtvRpodfVbnCK_E_LEMg7Z4TJG2O9I,14284
778
766
  wisent/core/evaluators/benchmark_specific/__init__.py,sha256=hTcHcv3GEQQWZ7Z_puSUbI8LlQuzljr_-GtMA19wiro,1432
779
767
  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
768
+ wisent/core/evaluators/benchmark_specific/apps_evaluator.py,sha256=_BoQjf6IouWWWMoiKk83-woZA7a73dBZC_qKKYEs9v0,4322
769
+ wisent/core/evaluators/benchmark_specific/conala_evaluator.py,sha256=z2tlm9mnMbglnnfEn2h_d9X1Z9D0KWSjFBKAB7oh0AY,6475
781
770
  wisent/core/evaluators/benchmark_specific/exact_match_evaluator.py,sha256=rc-QfiAYOzAkd5MAfghj7kQ6pEEYFtr4630pt-HRXmE,2756
782
771
  wisent/core/evaluators/benchmark_specific/f1_evaluator.py,sha256=6_9_U3mUKBSBpOFq7HGXnN3xrmGOSV-H2agTE_s1I_g,6608
783
772
  wisent/core/evaluators/benchmark_specific/generation_evaluator.py,sha256=6XNmp774qI9PqzybqQEwKgI_gn3ui3mIZc71GT5JXcY,20052
@@ -789,7 +778,7 @@ wisent/core/evaluators/benchmark_specific/polymath_evaluator.py,sha256=bZfGRp_ad
789
778
  wisent/core/evaluators/benchmark_specific/coding/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
790
779
  wisent/core/evaluators/benchmark_specific/coding/solution_generator.py,sha256=PX3RdwhrrYG_2nx10GVvQjdq-_6AFP-Q-yEGFA9rYpc,9329
791
780
  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
781
+ wisent/core/evaluators/benchmark_specific/coding/metrics/evaluator.py,sha256=KHlWIMkQcuKnES8H3qEo9IHMPAlsbhnF5yGCde-mUHQ,15313
793
782
  wisent/core/evaluators/benchmark_specific/coding/metrics/passk.py,sha256=WCkdt9a1KG2qoHNQtXdKPupVNQtppPyUGeq5yF6ZO7M,2059
794
783
  wisent/core/evaluators/benchmark_specific/coding/metrics/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
795
784
  wisent/core/evaluators/benchmark_specific/coding/metrics/core/atoms.py,sha256=I4wwOUTDjJ_4Jo6Bbpl8Tkv0Uz9zpwNjzeWBrXru-9g,1001
@@ -825,21 +814,21 @@ wisent/core/evaluators/custom/examples/__init__.py,sha256=Jho4mJgB6eGc5CaRjtcyrO
825
814
  wisent/core/evaluators/custom/examples/desklib_detector.py,sha256=1ysygPE-TpO7sKqOEZtOvdGMv5nLp6n57MawU7mH3Ic,5566
826
815
  wisent/core/evaluators/custom/examples/gptzero.py,sha256=El4Sypuee0a6QMv7HrPhfqfTeJbaEXQRLjuV_clPlgg,6129
827
816
  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
817
+ wisent/core/evaluators/custom/examples/humanization_coherent.py,sha256=4jad6VsPs9QSyDH0PKwhyjsxTvIM6a0tfZsOseeF46I,6628
829
818
  wisent/core/evaluators/custom/examples/roberta_detector.py,sha256=mq9MHdm6sosuzV8qRt-rj6gaYFaWsodgfBt1ZpCqwFM,6107
830
819
  wisent/core/evaluators/oracles/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
831
820
  wisent/core/evaluators/oracles/interactive.py,sha256=f3v2_N17fKzGyeOxONRJbrbn8i5uMeZmRvMmF0ShNf0,2638
832
821
  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
822
+ wisent/core/evaluators/oracles/truthfulqa_gen_evaluator.py,sha256=fUejtchm8bWFQitW-xIPkvo2BYvpyjGVFU-FR_s60EE,5954
834
823
  wisent/core/evaluators/oracles/user_specified.py,sha256=V1dKrNj3Oq7UC_I7DT0WGnktP7R_DSW6UAwDdrA8SnE,2360
835
824
  wisent/core/evaluators/personalization/__init__.py,sha256=t3lWkNFXA27eDzEQIZnkpasjGJ0w6UaZee4dJEQpFEQ,311
836
825
  wisent/core/evaluators/personalization/alignment.py,sha256=UcF5ismNElPbWkJNb-EuUKZraYE5aI46SAXoniQxQwE,5449
837
- wisent/core/evaluators/personalization/coherence.py,sha256=PM6_m9au2jntw6z1IjsNurl1oH8VfvPrwJZLJROXZiQ,11582
826
+ wisent/core/evaluators/personalization/coherence.py,sha256=f_0VHXm2GbS68JnQkYXS7tuR1sH6vsuLqSvjRB6M9kM,13641
838
827
  wisent/core/evaluators/personalization/difference.py,sha256=8D9azBdqciYo4IJt2lwrFGrwpm7q3Kyze6K_yBYPHoo,2649
839
828
  wisent/core/modalities/__init__.py,sha256=jtowcmTMQ_kOYai7FKwXY0vqSYQ8vews84VSMH73XFM,20138
840
829
  wisent/core/models/__init__.py,sha256=G9JfaVU_CXCG_wnz5D5gjDJ-_G7zqHFmjaixDpbCixM,456
841
830
  wisent/core/models/inference_config.py,sha256=0n8647avNb4w0o0UsbWnU_eh6cMU23ppdJ-FDYx8mAY,3704
842
- wisent/core/models/wisent_model.py,sha256=TvE1ZmSVCKwN5gs8c5HX06ABGsCw7G9RMzEer75ilvw,36518
831
+ wisent/core/models/wisent_model.py,sha256=ZOiT9WfkeZneqfiu_aS8XSIg18As204GR1j28R2D514,36526
843
832
  wisent/core/models/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
844
833
  wisent/core/models/core/atoms.py,sha256=7QMaJHl6axH6Iy0woZFsgEsBga1IaMhVq7q7JoASRic,16107
845
834
  wisent/core/opti/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -848,20 +837,21 @@ wisent/core/opti/core/atoms.py,sha256=7D6TxhEcBKxCplajICCfpZQ0g2rsKB6iWoRB0Mwr7y
848
837
  wisent/core/opti/methods/__init__.py,sha256=AJmvQZqWRAJ_EYzrF28uA1mN37C-AtVnUgCKfXaD0qQ,365
849
838
  wisent/core/opti/methods/opti_classificator.py,sha256=qVnpA57ncSsPLOby-rXvEOI_c_cRp06pJSsOLtO-tuw,6864
850
839
  wisent/core/opti/methods/opti_steering.py,sha256=Lr8dYKpqgGvuJauvYl7FNUMtR1cUAl93aepiqPIeslw,5407
851
- wisent/core/opti/methods/opti_weights.py,sha256=cfJziHJqd7P43sbZFDLqXmlyW-po-7yNmfKkCp6LBzc,19656
840
+ wisent/core/opti/methods/opti_weights.py,sha256=0dUN0lqMsRfB0Fcv851Kv-alseuS9N0Z4O1B12Y-qyU,21104
852
841
  wisent/core/optuna/__init__.py,sha256=efbr9QeNaeWMle9pppWBJXs5ObiPti9QZBQGBvuli-s,2250
853
842
  wisent/core/optuna/classifier/__init__.py,sha256=vv2wCAbw8ccZxq3vxrQt9plUdbr7VJj-t2rRnh6jBR8,819
854
- wisent/core/optuna/classifier/activation_generator.py,sha256=QB4Me1DlzmjN9C4En9tpfgV9K_qeO9Iqo7igTJw8LyA,14448
843
+ wisent/core/optuna/classifier/activation_generator.py,sha256=9zV4qEGVovBP_3HYQq0UXArZ_xPSIdjHEiGHzBveikY,14432
855
844
  wisent/core/optuna/classifier/classifier_cache.py,sha256=Q8PNBJ2rpHWvAQRVPR_ax0zvnrbRKN0yUjOhCwXf9jw,17439
856
845
  wisent/core/optuna/classifier/optuna_classifier_optimizer.py,sha256=T1K6CSWreQeX4X-AeAT0jaf5hBDp1bP92LN_tDzm500,25490
857
846
  wisent/core/optuna/steering/__init__.py,sha256=C1O9Y2AMFsFvLq6fmTW_wJ5vnRXKVGftykRw3NaiHwM,508
858
847
  wisent/core/optuna/steering/bigcode_evaluator_wrapper.py,sha256=ZVp159jbcdUAf8A2q5AwINak0n6YqBUuq0iDLvXObyQ,7157
859
848
  wisent/core/optuna/steering/data_utils.py,sha256=PUQ8Gqu8oPfoGGFi_lTfaPXhTZ5MRQO2Gf73vRa7NDU,12431
860
849
  wisent/core/optuna/steering/metrics.py,sha256=6UbW9AMrRpT1pgvDVK9rb1-ieqqOq1I1KdVh6MtB4CI,16394
861
- wisent/core/optuna/steering/steering_optimization.py,sha256=w-wt0AUNBr5RmYvg58pepx2KDqQAjlM9FBU2a2OgaRk,44893
850
+ wisent/core/optuna/steering/steering_optimization.py,sha256=V5iu5uttoT1Uq6Lj5jiUfnJ9-vIXVyAVD6UzXHdvXWA,45016
862
851
  wisent/core/parser_arguments/__init__.py,sha256=MpBsnsPcZQ8cqAn1tN4OYfb3oMQraNW-23g3Z5gSqeg,299
863
852
  wisent/core/parser_arguments/agent_parser.py,sha256=PG3V-F9xKE53guiiQ2GudGKZnjvtoAbdyMI8Qhm_sZw,4971
864
853
  wisent/core/parser_arguments/check_linearity_parser.py,sha256=HnYdSYOOP9MvVyRjO2ZvTIlKU-y_Hw3yXL5F_ehjZXs,2013
854
+ wisent/core/parser_arguments/cluster_benchmarks_parser.py,sha256=QvEjOIRZswffCNMcEsIgCt1a7V-CzaHbFVWTaDF2-7U,889
865
855
  wisent/core/parser_arguments/configure_model_parser.py,sha256=LmsS0mYF7eP-zpYB7SdEbtFZLVtPtcbkEi9CMkYqMVo,350
866
856
  wisent/core/parser_arguments/create_steering_vector_parser.py,sha256=GvyHt0fqwVk5dDmAsilOttsybZ3jj5Z6dQWzE8EUj-Y,1735
867
857
  wisent/core/parser_arguments/diagnose_pairs_parser.py,sha256=i_kV0sEYkpWEur0BK32Y3nMMRwjz7bAjUY9BbhqSDws,591
@@ -874,11 +864,11 @@ wisent/core/parser_arguments/generate_pairs_from_task_parser.py,sha256=wEZLIXocq
874
864
  wisent/core/parser_arguments/generate_pairs_parser.py,sha256=nDTQ63bvjl-GcSrMdtcR_Pdh1dzbqtRnGaSCFhPO-4Q,1918
875
865
  wisent/core/parser_arguments/generate_responses_parser.py,sha256=LnjIx7gOdysEMGwUIHFHlaEOE306qr4rXto9gzBSjkg,1345
876
866
  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
867
+ wisent/core/parser_arguments/generate_vector_from_task_parser.py,sha256=88gnq8_fmQAmn3zJR1q77rIBFiOBz21nfyJQmd_H_P8,4873
878
868
  wisent/core/parser_arguments/generate_vector_parser.py,sha256=bf2PHAdvNZetXaSe3sKarvK6McIODPdmRSNRQN_rHto,3622
879
869
  wisent/core/parser_arguments/get_activations_parser.py,sha256=pCBET7CsFz-8WFE3mb677o4W-vVa09wz-3YDMx-8ukA,2445
880
870
  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
871
+ wisent/core/parser_arguments/main_parser.py,sha256=BKQOWO-6l3USgiJrldH_tZ9x-io2grjK-UaYw615hQI,12345
882
872
  wisent/core/parser_arguments/model_config_parser.py,sha256=RlcSyQkj0uDlmJJE8sMq_CjvJ1VdCcjf8_BqspsSHd0,3210
883
873
  wisent/core/parser_arguments/modify_weights_parser.py,sha256=077Gz6GLk1aR2k6rboU4tcVBgxHaTB01sSRx7ZqlzPU,9336
884
874
  wisent/core/parser_arguments/monitor_parser.py,sha256=qo3vyyVmdZBkhGuhlHqbEzTiIKcdFIvG3IIuc-DIRQE,1234
@@ -888,10 +878,10 @@ wisent/core/parser_arguments/optimization_cache_parser.py,sha256=R54XzVY4GUgpXu8
888
878
  wisent/core/parser_arguments/optimize_classification_parser.py,sha256=tzHAvOPIbXP8risT6eMICpsRfO6RETwP4s8qIAe60ms,4170
889
879
  wisent/core/parser_arguments/optimize_parser.py,sha256=HTuhFspCAZkNcYhP69-JmaT91CRm_wV9IaF-BFvg_I8,4603
890
880
  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
881
+ wisent/core/parser_arguments/optimize_steering_parser.py,sha256=fRG64w-eb0ezUp9T5-zsMJ8X8L4Z6C9myeg21R8DTuI,28731
882
+ wisent/core/parser_arguments/optimize_weights_parser.py,sha256=7SEOhgBs9GWGLyPzNfZHQztt4o24HHUPgCf7sbW1kVo,13904
893
883
  wisent/core/parser_arguments/synthetic_parser.py,sha256=99fGg_k9vQl80COuVZw5LOYLRQ0sDlZ4ozwsx0T--Qs,5135
894
- wisent/core/parser_arguments/tasks_parser.py,sha256=EUoRxKy3I5ekv7FGORAGLMeECBg2QL_fIv8nurI9jsQ,22992
884
+ wisent/core/parser_arguments/tasks_parser.py,sha256=MNSijAG6zzinuwy9lIOwJxrL0JavM7MKx443Nwe7uEs,22440
895
885
  wisent/core/parser_arguments/train_unified_goodness_parser.py,sha256=5BICCLDk6UrUjDXiwsOIYJYx9ubPwwPhC70gAP45Rq0,4962
896
886
  wisent/core/parser_arguments/utils.py,sha256=XTtssIt05HlAyO6kMPZvu4l_UIZSMDaV-gl50NO7qOw,3543
897
887
  wisent/core/prompts/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -905,15 +895,16 @@ wisent/core/prompts/prompt_stratiegies/multiple_choice.py,sha256=L4dOtocXte91Mn2
905
895
  wisent/core/prompts/prompt_stratiegies/role_playing.py,sha256=Bl_4A-OJgepDB33iRLB-fTdCLmy4QvfGKJ7kIq1ApUE,1101
906
896
  wisent/core/steering_methods/__init__.py,sha256=ULxu5kHL5jCyrGoPAUiOYtuxU_iTYYY7e0frABdflfk,1364
907
897
  wisent/core/steering_methods/preflight.py,sha256=26Pxx7b3TBRSHfnKF2NWhYN2mjTCnbnd3k7advIqjfU,12009
908
- wisent/core/steering_methods/registry.py,sha256=kD76e_gtK-LwdPO6qz1Ai-dH_jYSobfOVSY5qzaMeUg,22482
898
+ wisent/core/steering_methods/registry.py,sha256=t7gvgCa294HhY__Nb_swt1TvfEWNRXqh578LJxNuWWc,24066
909
899
  wisent/core/steering_methods/rotator.py,sha256=Lt_IW2I1u9h79KLyiTtLIocGhdW0DxcQylMyqC8gClo,4053
910
900
  wisent/core/steering_methods/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
911
- wisent/core/steering_methods/core/atoms.py,sha256=Dtzg1t6nVAOJYRvQLptQchyjhiSx8mEgv2ImtXNCT_s,5485
901
+ wisent/core/steering_methods/core/atoms.py,sha256=kDN-lvaQnMi9UTCdz-4rZ8x5baqrsVGVYGO-P2EpVUY,5377
912
902
  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
903
+ wisent/core/steering_methods/methods/caa.py,sha256=EsodLtBV66YR7ROngU7-9fJFhzWcmzo7HFjZCNE28Ko,1717
904
+ wisent/core/steering_methods/methods/hyperplane.py,sha256=ST44J2fspz1sFkda-w5dAJSE4SHRG4ujfcnA0PKnbGo,2919
905
+ wisent/core/steering_methods/methods/prism.py,sha256=Cb18ip7tl57hz7KW6br4EsLWxYP5__bMYppsyBEPzIY,23995
906
+ wisent/core/steering_methods/methods/pulse.py,sha256=aowQtMONcBkNaO99gT7qVClifKyyEdWPQqKMv-PwjSE,25801
907
+ wisent/core/steering_methods/methods/titan.py,sha256=OVj02CrOlO_YVKxiOK6kBZUiViLZL3ot2709J91tFf8,41521
917
908
  wisent/core/synthetic/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
918
909
  wisent/core/synthetic/cleaners/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
919
910
  wisent/core/synthetic/cleaners/deduper_cleaner.py,sha256=XI19V8IpX5lGWasW43Qa2peEEcxsrAfmB_W9uKzw11U,1571
@@ -955,7 +946,7 @@ wisent/core/tracking/__init__.py,sha256=rda2gaV_tFP6CChzx8TBQUrh_PObqvX35-BHJCxi
955
946
  wisent/core/tracking/latency.py,sha256=uKibYRodOurQoCh5MqfJwo_k5Ho4H5dFQ0p_nddHGOg,21909
956
947
  wisent/core/tracking/memory.py,sha256=Fwg92c_M997uToNLIZvFlHbp3F5TTrUZsA_X25J-QfM,13201
957
948
  wisent/core/trainers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
958
- wisent/core/trainers/steering_trainer.py,sha256=19GNV7QX9mBl5awPwXANK6haW-Xf4vCyYKgdAjvvH0o,15759
949
+ wisent/core/trainers/steering_trainer.py,sha256=yJ44B5l0jVHeWKQ2yajSGjKvBLSEIUfdUo5SM8Yt_ow,15105
959
950
  wisent/core/trainers/core/__init__.py,sha256=zkv6f6Auk4dbJ5PTQP96gP2TUM55W3tc1flHLWBvI4M,1246
960
951
  wisent/core/trainers/core/atoms.py,sha256=ycWk0G-7EIAEOnQL-o5_V5B8KTQ7CQUilGF4ibjighM,1536
961
952
  wisent/core/utils/__init__.py,sha256=ee4tMv5W_qxtTwRS-n8ZCl69CgBUqPGjugrkomOhzhE,1419
@@ -1032,632 +1023,6 @@ wisent/examples/scripts/2/test_babilong_pairs.json,sha256=OOWUdyqtMXUv2_WWIqTTEY
1032
1023
  wisent/examples/scripts/2/test_bangla_mmlu_evaluation.json,sha256=OtR2JvARwR99QphQDprHiyv7wjLPK9583gR_rCM9RQU,2073
1033
1024
  wisent/examples/scripts/2/test_bangla_mmlu_pairs.json,sha256=GR9eXDwYdp0CBOyOZ9XF_yHv07KlDcpv8BpPF5wovXw,1340
1034
1025
  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
1026
  wisent/parameters/__init__.py,sha256=dwyW8F6LSMPM-H4zQX6gn7x5qPs78oNAMYRiRPJ8UCk,36
1662
1027
  wisent/parameters/lm_eval/all_lm_eval_task_families.json,sha256=MAFcazf3UoDplgaeWdkcKpAOzFRHyc5yLoXLqOwyFvc,2555
1663
1028
  wisent/parameters/lm_eval/broken_in_lm_eval.json,sha256=C8vWqqJBRSUnLwCnuxwt_hfpaykfkxXN8NRJUCUixzs,120
@@ -1671,18 +1036,21 @@ wisent/parameters/lm_eval/no_readmes.json,sha256=T1PNoYwrqgwDVLtfmj7L5e0Sq02OEbq
1671
1036
  wisent/parameters/lm_eval/not_lm_eval_tasks.json,sha256=zcdUnstl_OA52DO5Sf8ZAPkNcu9B-bfuyTXxPm8MEr8,1964
1672
1037
  wisent/parameters/lm_eval/read_tasks.json,sha256=MLOSbTPLXhMcPVw5QJosIIlUiRTFeQmbFZ16wFcMMKk,3401
1673
1038
  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
1039
+ wisent/parameters/lm_eval/track_progress_not_lm_eval_tasks.json,sha256=6dujwX_RNY1MZFbWMb1MtXYK0CdN1QTC6ppPWpxO_64,3472
1675
1040
  wisent/parameters/tasks/missing_task_families.json,sha256=h6g_2CYNQIW5SqRdcymDe4miWkGkVoD3yR0eOFZH4aw,65107
1676
1041
  wisent/parameters/tasks/remaining_tasks_to_implement.json,sha256=8_YDVQ439DdvC3daNW7v8y0Zth1eUxum6QzdnKDp6a8,3432
1677
1042
  wisent/parameters/tasks/risks.json,sha256=8U1LAY6Zrv68naYMzVlOrGgTBNL0l-9nC7moa-ljSHI,136
1678
1043
  wisent/parameters/tasks/skills.json,sha256=63KirwoeFTabMsQ3B7IdNyQLyl8857J0gJF02tTNPQo,215
1679
1044
  wisent/parameters/tasks/tasks.json,sha256=7HvSgYjO5dQfHfIwLB3D0h0lVQqpf2tKnU6U36j-r6Y,1140923
1680
- wisent/scripts/run_quality_metrics_sweep.sh,sha256=NFqvIc1JB5fYYDXFBQeulQZV2cavbmsRQ2QWBU9aM10,10733
1045
+ wisent/scripts/run_quality_metrics_sweep.sh,sha256=Lj8euHTGBO3QMdNhpW5zFXWPd4bPPYLJRZmZPhQvhiM,10493
1681
1046
  wisent/tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1047
+ wisent/tests/test_aggregation_geometry.py,sha256=hh8aJ6VyuN_3xA0x9YaoPhu_lMVYKupUxpVyFtEF_18,9277
1682
1048
  wisent/tests/test_all_cli_commands.py,sha256=7KqSZ8FsJEwXbgp0w_cfVtv9jgd-9P52fH3nx3jGF9k,23432
1049
+ wisent/tests/test_detector_accuracy.py,sha256=HnmBBwCMZC0bM738o8oCdk4dZmlA8Sr228iLdAbTNio,6484
1683
1050
  wisent/tests/test_geometry_comprehensive.py,sha256=lUyBfZYudMMaX2-nV2QhtQcchbBaz8bLi6fqmMRQOig,13693
1051
+ wisent/tests/test_geometry_exhaustive.py,sha256=yQnUIjNTW8DaC_bwU_rfHThOutLoMGxTHxsbkbqCZNQ,49679
1684
1052
  wisent/tests/test_titan_geometry.py,sha256=lKWsFSCGEHFHD-CW_lz0vfATb6iFij0lrrLosBQJmkc,9710
1685
- wisent/tests/visualize_geometry.py,sha256=mmv9bjA7MmjSZ1wS-J2h8G4g9YqtReOPmOE94giCibE,6119
1053
+ wisent/tests/visualize_geometry.py,sha256=dhwyRlqjZfIY9xxJO5rFjLOfdmPyjhg9sjZTcd4qBIY,13861
1686
1054
  wisent/tests/examples/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1687
1055
  wisent/tests/examples/cli/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1688
1056
  wisent/tests/examples/cli/activations/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -1712,9 +1080,9 @@ wisent/tests/nosense/__init__.py,sha256=sH3x4jRPzFM3YmQkdrwJoz-BdOQ1Bh6F95G5HWyI
1712
1080
  wisent/tests/nosense/base_nosense.py,sha256=a18dBv1378nHly7OCIuk-bCcLnubss3XXDC1ex0zCK8,2633
1713
1081
  wisent/tests/nosense/math500_nosense.py,sha256=My0dHsr4OFOiTxb_VDKmGzpoMyzAtqXlHhA0oPfaG7s,2389
1714
1082
  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,,
1083
+ wisent-0.7.701.dist-info/licenses/LICENSE,sha256=wy0iaw8b2tyqZAfKHib3lP3PJ9o88FDCg92oUHh3sDQ,1073
1084
+ wisent-0.7.701.dist-info/METADATA,sha256=ZO6yC67KziyCpX9sXltKYVciAxX2s5jMm9AYTmyRWt0,2125
1085
+ wisent-0.7.701.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
1086
+ wisent-0.7.701.dist-info/entry_points.txt,sha256=BM76j3xjtIcVZGk24iDf5w18s6SuqeOpaiAxfZhpnY8,49
1087
+ wisent-0.7.701.dist-info/top_level.txt,sha256=2Ts9Iyldnb3auIN2HBBaHPknRy7nSRDm2f6RGzYgr8A,7
1088
+ wisent-0.7.701.dist-info/RECORD,,