wisent 0.7.379__py3-none-any.whl → 0.7.701__py3-none-any.whl

wisent/examples/scripts/results/benchmark_tags.json

@@ -1,917 +0,0 @@
-{
-  "aclue": {
-    "tags": ["multilingual", "history"]
-  },
-  "acpbench": {
-    "tags": ["reasoning"]
-  },
-  "aexams": {
-    "tags": ["multilingual", "science"]
-  },
-  "afrimgsm": {
-    "tags": ["mathematics", "multilingual"]
-  },
-  "afrimmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "afrixnli": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "afrobench": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "afrobench_adr": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "afrobench_afriqa": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "afrobench_afrisenti": {
-    "tags": ["multilingual", "bias"]
-  },
-  "afrobench_belebele": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "afrobench_flores": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "afrobench_injongointent": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "afrobench_mafand": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "afrobench_masakhaner": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "afrobench_masakhanews": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "afrobench_masakhapos": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "afrobench_naijarc": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "afrobench_nollysenti": {
-    "tags": ["multilingual", "bias"]
-  },
-  "afrobench_ntrex": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "afrobench_openai_mmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "afrobench_salt": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "afrobench_sib": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "afrobench_uhura-arc-easy": {
-    "tags": ["multilingual", "science", "hallucination"]
-  },
-  "afrobench_xlsum": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "agieval": {
-    "tags": ["general knowledge", "reasoning"]
-  },
-  "alghafa_copa_ar": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "alghafa_piqa_ar": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "anli": {
-    "tags": ["reasoning", "adversarial robustness"]
-  },
-  "arab_culture": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "arab_culture_completion": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "arabic_leaderboard_complete": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "arabic_leaderboard_light": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "arabicmmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "aradice": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "arc": {
-    "tags": ["science", "reasoning"]
-  },
-  "arc_mt": {
-    "tags": ["multilingual", "science"]
-  },
-  "arithmetic": {
-    "tags": ["mathematics"]
-  },
-  "asdiv": {
-    "tags": ["mathematics", "reasoning"]
-  },
-  "babi": {
-    "tags": ["reasoning"]
-  },
-  "basque_bench": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "basqueglue": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "bbh": {
-    "tags": ["reasoning", "adversarial robustness"]
-  },
-  "bbq": {
-    "tags": ["bias"]
-  },
-  "belebele": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "benchmarks_multimedqa": {
-    "tags": ["medical", "general knowledge"]
-  },
-  "bertaqa": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "bigbench": {
-    "tags": ["reasoning", "general knowledge"]
-  },
-  "blimp": {
-    "tags": ["reasoning"]
-  },
-  "c4": {
-    "tags": ["general knowledge"]
-  },
-  "careqa": {
-    "tags": ["medical", "multilingual"]
-  },
-  "catalan_bench": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "ceval": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "chartqa": {
-    "tags": ["reasoning"]
-  },
-  "cmmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "code_x_glue": {
-    "tags": ["coding"]
-  },
-  "commonsense_qa": {
-    "tags": ["reasoning", "general knowledge"]
-  },
-  "copal_id": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "coqa": {
-    "tags": ["reasoning"]
-  },
-  "crows_pairs": {
-    "tags": ["bias"]
-  },
-  "csatqa": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "darija_bench": {
-    "tags": ["multilingual"]
-  },
-  "darija_bench_darija_sentiment": {
-    "tags": ["multilingual", "bias"]
-  },
-  "darija_bench_darija_summarization": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "darija_bench_darija_translation": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "darija_bench_darija_transliteration": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "darijahellaswag": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "darijammlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "drop": {
-    "tags": ["reasoning", "mathematics"]
-  },
-  "egyhellaswag": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "egymmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "eq_bench": {
-    "tags": ["reasoning"]
-  },
-  "eus_exams": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "eus_proficiency": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "eus_reading": {
-    "tags": ["multilingual", "reasoning", "long context"]
-  },
-  "eus_trivia": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "evalita_llm": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "fda": {
-    "tags": ["reasoning"]
-  },
-  "fld": {
-    "tags": ["reasoning"]
-  },
-  "french_bench": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "galician_bench": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "glianorex": {
-    "tags": ["medical", "hallucination"]
-  },
-  "global_mmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "glue": {
-    "tags": ["reasoning"]
-  },
-  "gpqa": {
-    "tags": ["science", "reasoning"]
-  },
-  "groundcocoa": {
-    "tags": ["reasoning"]
-  },
-  "gsm_plus": {
-    "tags": ["mathematics", "adversarial robustness"]
-  },
-  "gsm8k": {
-    "tags": ["mathematics", "reasoning"]
-  },
-  "gsm8k_platinum": {
-    "tags": ["mathematics", "reasoning"]
-  },
-  "haerae": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "headqa": {
-    "tags": ["medical", "multilingual"]
-  },
-  "hellaswag": {
-    "tags": ["reasoning"]
-  },
-  "hendrycks_ethics": {
-    "tags": ["bias"]
-  },
-  "hendrycks_math": {
-    "tags": ["mathematics", "reasoning"]
-  },
-  "histoires_morales": {
-    "tags": ["multilingual", "bias"]
-  },
-  "hrm8k": {
-    "tags": ["mathematics", "multilingual"]
-  },
-  "humaneval": {
-    "tags": ["coding"]
-  },
-  "ifeval": {
-    "tags": ["reasoning"]
-  },
-  "inverse_scaling": {
-    "tags": ["adversarial robustness"]
-  },
-  "japanese_leaderboard": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "jsonschema_bench": {
-    "tags": ["coding"]
-  },
-  "kbl": {
-    "tags": ["law", "multilingual"]
-  },
-  "kmmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "kobest": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "kormedmcqa": {
-    "tags": ["medical", "multilingual"]
-  },
-  "lambada": {
-    "tags": ["reasoning", "long context"]
-  },
-  "lambada_cloze": {
-    "tags": ["reasoning", "long context"]
-  },
-  "lambada_multilingual": {
-    "tags": ["multilingual", "reasoning", "long context"]
-  },
-  "lambada_multilingual_stablelm": {
-    "tags": ["multilingual", "reasoning", "long context"]
-  },
-  "leaderboard": {
-    "tags": ["reasoning", "general knowledge"]
-  },
-  "libra": {
-    "tags": ["multilingual", "long context"]
-  },
-  "lingoly": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "logiqa": {
-    "tags": ["reasoning"]
-  },
-  "logiqa2": {
-    "tags": ["reasoning"]
-  },
-  "longbench": {
-    "tags": ["long context", "reasoning"]
-  },
-  "mastermind": {
-    "tags": ["reasoning"]
-  },
-  "mathqa": {
-    "tags": ["mathematics", "reasoning"]
-  },
-  "mbpp": {
-    "tags": ["coding"]
-  },
-  "mc_taco": {
-    "tags": ["reasoning"]
-  },
-  "med_concepts_qa": {
-    "tags": ["medical"]
-  },
-  "meddialog": {
-    "tags": ["medical"]
-  },
-  "mediqa_qa2019": {
-    "tags": ["medical"]
-  },
-  "medmcqa": {
-    "tags": ["medical"]
-  },
-  "medqa": {
-    "tags": ["medical"]
-  },
-  "medtext": {
-    "tags": ["medical"]
-  },
-  "mela": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "meqsum": {
-    "tags": ["medical", "creative writing"]
-  },
-  "metabench": {
-    "tags": ["reasoning", "general knowledge"]
-  },
-  "mgsm": {
-    "tags": ["mathematics", "multilingual"]
-  },
-  "mimic_repsum": {
-    "tags": ["medical", "creative writing"]
-  },
-  "minerva_math": {
-    "tags": ["mathematics", "reasoning"]
-  },
-  "mlqa": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "mmlu": {
-    "tags": ["general knowledge", "reasoning"]
-  },
-  "mmlu_pro": {
-    "tags": ["general knowledge", "reasoning"]
-  },
-  "mmlu_prox": {
-    "tags": ["multilingual", "general knowledge", "reasoning"]
-  },
-  "mmlu-pro-plus": {
-    "tags": ["general knowledge", "reasoning", "adversarial robustness"]
-  },
-  "mmlusr": {
-    "tags": ["general knowledge", "adversarial robustness"]
-  },
-  "mmmu": {
-    "tags": ["general knowledge", "reasoning"]
-  },
-  "model_written_evals": {
-    "tags": ["bias", "sycophancy"]
-  },
-  "moral_stories": {
-    "tags": ["bias"]
-  },
-  "mts_dialog": {
-    "tags": ["medical"]
-  },
-  "multiblimp": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "mutual": {
-    "tags": ["reasoning"]
-  },
-  "noreval": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "noreval_ask_gec": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "noticia": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "nq_open": {
-    "tags": ["general knowledge", "reasoning"]
-  },
-  "okapi_arc_multilingual": {
-    "tags": ["multilingual", "science", "reasoning"]
-  },
-  "okapi_hellaswag_multilingual": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "okapi_mmlu_multilingual": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "okapi_truthfulqa_multilingual": {
-    "tags": ["multilingual", "hallucination"]
-  },
-  "olaph": {
-    "tags": ["medical"]
-  },
-  "openbookqa": {
-    "tags": ["science", "reasoning"]
-  },
-  "paloma": {
-    "tags": ["general knowledge"]
-  },
-  "paws-x": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "pile": {
-    "tags": ["general knowledge"]
-  },
-  "pile_10k": {
-    "tags": ["general knowledge"]
-  },
-  "piqa": {
-    "tags": ["reasoning", "science"]
-  },
-  "polemo2": {
-    "tags": ["multilingual", "bias"]
-  },
-  "portuguese_bench": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "prost": {
-    "tags": ["reasoning", "science"]
-  },
-  "pubmedqa": {
-    "tags": ["medical"]
-  },
-  "qa4mre": {
-    "tags": ["reasoning"]
-  },
-  "qasper": {
-    "tags": ["science", "reasoning"]
-  },
-  "race": {
-    "tags": ["reasoning"]
-  },
-  "realtoxicityprompts": {
-    "tags": ["toxicity"]
-  },
-  "ruler": {
-    "tags": ["long context", "reasoning"]
-  },
-  "sciq": {
-    "tags": ["science"]
-  },
-  "score": {
-    "tags": ["adversarial robustness"]
-  },
-  "scrolls": {
-    "tags": ["long context", "reasoning"]
-  },
-  "simple_cooccurrence_bias": {
-    "tags": ["bias"]
-  },
-  "siqa": {
-    "tags": ["reasoning"]
-  },
-  "spanish_bench": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "squad_completion": {
-    "tags": ["reasoning"]
-  },
-  "squadv2": {
-    "tags": ["reasoning"]
-  },
-  "storycloze": {
-    "tags": ["reasoning"]
-  },
-  "super_glue": {
-    "tags": ["reasoning"]
-  },
-  "swag": {
-    "tags": ["reasoning"]
-  },
-  "swde": {
-    "tags": ["reasoning"]
-  },
-  "tinyBenchmarks": {
-    "tags": ["reasoning", "general knowledge"]
-  },
-  "tmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "tmmluplus": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "toxigen": {
-    "tags": ["toxicity"]
-  },
-  "translation": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "triviaqa": {
-    "tags": ["general knowledge", "reasoning"]
-  },
-  "truthfulqa": {
-    "tags": ["hallucination"]
-  },
-  "truthfulqa-multi": {
-    "tags": ["multilingual", "hallucination"]
-  },
-  "turkishmmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "unscramble": {
-    "tags": ["reasoning"]
-  },
-  "webqs": {
-    "tags": ["general knowledge"]
-  },
-  "wikitext": {
-    "tags": ["general knowledge"]
-  },
-  "winogender": {
-    "tags": ["bias"]
-  },
-  "winogrande": {
-    "tags": ["reasoning"]
-  },
-  "wmdp": {
-    "tags": ["science"]
-  },
-  "wmt2016": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "wsc273": {
-    "tags": ["reasoning"]
-  },
-  "xcopa": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "xnli": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "xnli_eu": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "xquad": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "xstorycloze": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "xwinograd": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "20_newsgroups": {
-    "tags": ["reasoning"]
-  },
-  "AraDiCE": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "ArabCulture": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "acp_bench": {
-    "tags": ["reasoning"]
-  },
-  "acp_bench_hard": {
-    "tags": ["reasoning"]
-  },
-  "afrimgsm_direct_amh": {
-    "tags": ["mathematics", "multilingual"]
-  },
-  "afrimmlu_direct_amh": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "afrixnli_en_direct_amh": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "ag_news": {
-    "tags": ["reasoning"]
-  },
-  "aime": {
-    "tags": ["mathematics"]
-  },
-  "aime2024": {
-    "tags": ["mathematics"]
-  },
-  "aime2025": {
-    "tags": ["mathematics"]
-  },
-  "apps": {
-    "tags": ["coding"]
-  },
-  "arabic_exams": {
-    "tags": ["multilingual", "science"]
-  },
-  "argument_topic": {
-    "tags": ["reasoning"]
-  },
-  "atis": {
-    "tags": ["reasoning"]
-  },
-  "babilong": {
-    "tags": ["long context", "reasoning"]
-  },
-  "bangla_mmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "banking77": {
-    "tags": ["reasoning"]
-  },
-  "basque-glue": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "bec2016eu": {
-    "tags": ["multilingual", "bias"]
-  },
-  "bhtc_v2": {
-    "tags": ["reasoning"]
-  },
-  "boolq": {
-    "tags": ["reasoning"]
-  },
-  "boolq-seq2seq": {
-    "tags": ["reasoning"]
-  },
-  "cb": {
-    "tags": ["reasoning"]
-  },
-  "chain_of_thought": {
-    "tags": ["reasoning"]
-  },
-  "claim_stance_topic": {
-    "tags": ["reasoning"]
-  },
-  "cnn_dailymail": {
-    "tags": ["creative writing"]
-  },
-  "codexglue_code_to_text_go": {
-    "tags": ["coding"]
-  },
-  "codexglue_code_to_text_java": {
-    "tags": ["coding"]
-  },
-  "codexglue_code_to_text_javascript": {
-    "tags": ["coding"]
-  },
-  "codexglue_code_to_text_php": {
-    "tags": ["coding"]
-  },
-  "codexglue_code_to_text_python": {
-    "tags": ["coding"]
-  },
-  "codexglue_code_to_text_ruby": {
-    "tags": ["coding"]
-  },
-  "coedit_gec": {
-    "tags": ["creative writing"]
-  },
-  "conala": {
-    "tags": ["coding"]
-  },
-  "concode": {
-    "tags": ["coding"]
-  },
-  "copa": {
-    "tags": ["reasoning"]
-  },
-  "dbpedia_14": {
-    "tags": ["reasoning"]
-  },
-  "doc_vqa": {
-    "tags": ["reasoning"]
-  },
-  "ds1000": {
-    "tags": ["coding"]
-  },
-  "ethos_binary": {
-    "tags": ["harmfulness"]
-  },
-  "evalita-mp": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "evalita-sp_sum_task_fp-small_p1": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "evalita_LLM": {
-    "tags": ["multilingual", "reasoning"]
-  },
-  "financial_tweets": {
-    "tags": ["reasoning"]
-  },
-  "flores": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "freebase": {
-    "tags": ["general knowledge", "reasoning"]
-  },
-  "global_mmlu_ar": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "gpt3_translation_benchmarks": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "hmmt": {
-    "tags": ["mathematics"]
-  },
-  "hmmt_feb_2025": {
-    "tags": ["mathematics"]
-  },
-  "humaneval_64_instruct": {
-    "tags": ["coding"]
-  },
-  "humaneval_instruct": {
-    "tags": ["coding"]
-  },
-  "humanevalpack": {
-    "tags": ["coding"]
-  },
-  "instruct_humaneval": {
-    "tags": ["coding"]
-  },
-  "instructhumaneval": {
-    "tags": ["coding"]
-  },
-  "iwslt2017-ar-en": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "iwslt2017-en-ar": {
-    "tags": ["multilingual", "creative writing"]
-  },
-  "law_stack_exchange": {
-    "tags": ["law", "reasoning"]
-  },
-  "ledgar": {
-    "tags": ["law", "reasoning"]
-  },
-  "livecodebench": {
-    "tags": ["coding"]
-  },
-  "livemathbench_cnmo_en": {
-    "tags": ["mathematics"]
-  },
-  "livemathbench_cnmo_zh": {
-    "tags": ["mathematics", "multilingual"]
-  },
-  "logieval": {
-    "tags": ["reasoning"]
-  },
-  "m_mmlu": {
-    "tags": ["multilingual", "general knowledge"]
-  },
-  "math": {
-    "tags": ["mathematics"]
-  },
-  "math500": {
-    "tags": ["mathematics"]
-  },
-  "mbpp_plus": {
-    "tags": ["coding"]
-  },
-  "medical_abstracts": {
-    "tags": ["medical", "reasoning"]
-  },
-  "mnli": {
-    "tags": ["reasoning"]
-  },
-  "multirc": {
-    "tags": ["reasoning"]
-  },
-  "ner": {
-    "tags": ["reasoning"]
-  },
-  "openllm": {
-    "tags": ["general knowledge", "reasoning"]
-  },
-  "pile_10k": {
-    "tags": ["general knowledge"]
-  },
-  "piqa": {
-    "tags": ["reasoning"]
-  },
-  "qasper": {
-    "tags": ["science", "reasoning"]
-  },
-  "mercury": {
-    "tags": ["science", "reasoning"]
-  },
-  "mocha": {
-    "tags": ["reasoning"]
-  },
-  "musr": {
-    "tags": ["reasoning"]
-  },
-  "openbookqa_fact": {
-    "tags": ["science", "reasoning"]
-  },
-  "pgeval": {
-    "tags": ["coding"]
-  },
-  "plurals": {
-    "tags": ["reasoning"]
-  },
-  "presencemindqa": {
-    "tags": ["reasoning"]
-  },
-  "pubmed_entailment": {
-    "tags": ["medical", "reasoning"]
-  },
-  "qasc": {
-    "tags": ["science", "reasoning"]
-  },
-  "quail": {
-    "tags": ["reasoning"]
-  },
-  "quartz": {
-    "tags": ["science", "reasoning"]
-  },
-  "quoref": {
-    "tags": ["reasoning"]
-  },
-  "reclor": {
-    "tags": ["reasoning"]
-  },
-  "scitail": {
-    "tags": ["science", "reasoning"]
-  },
-  "simpleqa": {
-    "tags": ["general knowledge", "hallucination"]
-  },
-  "socialiqa": {
-    "tags": ["reasoning"]
-  },
-  "stsb": {
-    "tags": ["reasoning"]
-  },
-  "sycophancy_eval": {
-    "tags": ["sycophancy"]
-  },
-  "tinygsm": {
-    "tags": ["mathematics"]
-  },
-  "webglm": {
-    "tags": ["reasoning"]
-  },
-  "wildbench": {
-    "tags": ["reasoning"]
-  },
-  "wildguard": {
-    "tags": ["harmfulness"]
-  },
-  "humaneval_plus": {
-    "tags": ["coding"]
-  },
-  "multiple": {
-    "tags": ["coding", "multilingual"]
-  },
-  "recode": {
-    "tags": ["coding"]
-  },
-  "polymath": {
-    "tags": ["mathematics"]
-  }
-}