| { |
| "manifest_version": "0.1", |
| "pinned_on": "2026-09-26", |
| "note": "Revisions, file paths, column names and label vocabularies below were read live from the Hugging Face dataset API and datasets-server on 2026-09-26. Row counts are intentionally absent: download.py measures them.", |
| "sources": [ |
| { |
| "key": "medmcqa", |
| "hf_repo": "openlifescienceai/medmcqa", |
| "revision": "91c6572c454088bf71b679ad90aa8dffcd0d5868", |
| "files": ["data/validation-00000-of-00001.parquet"], |
| "format": "parquet", |
| "license": "apache-2.0", |
| "redistributable": true, |
| "exposure": "unknown", |
| "tier": "mcq", |
| "task_type": "choice", |
| "n_options": 4, |
| "columns": ["id", "question", "opa", "opb", "opc", "opd", "cop", "choice_type", "exp", "subject_name", "topic_name"], |
| "model_visible": ["question", "opa", "opb", "opc", "opd"], |
| "gold_field": "cop", |
| "gold_vocab": ["a", "b", "c", "d"], |
| "hidden": ["cop", "exp", "subject_name", "topic_name"], |
| "stratify_by": "subject_name", |
| "target_n": 500, |
| "notes": "Validation split is the labelled eval split. The test split ships unlabelled gold in this repo, so do not use it. exp is a rationale and must never enter state." |
| }, |
| { |
| "key": "medqa_usmle", |
| "hf_repo": "GBaker/MedQA-USMLE-4-options", |
| "revision": "0fb93dd23a7339b6dcd27e241cb9b5eca62d4d18", |
| "files": ["phrases_no_exclude_test.jsonl"], |
| "format": "jsonl", |
| "license": "cc-by-4.0", |
| "redistributable": true, |
| "exposure": "likely", |
| "tier": "mcq", |
| "task_type": "choice", |
| "n_options": 4, |
| "columns": ["question", "answer", "options", "meta_info", "answer_idx", "metamap_phrases"], |
| "model_visible": ["question", "options"], |
| "gold_field": "answer_idx", |
| "gold_vocab": ["A", "B", "C", "D"], |
| "hidden": ["answer", "answer_idx", "metamap_phrases", "meta_info"], |
| "stratify_by": "meta_info", |
| "target_n": 500, |
| "notes": "options is a dict keyed A-D. answer is the gold option TEXT and answer_idx the gold letter: both are gold, both stay hidden. metamap_phrases is derived annotation, not part of the exam stem." |
| }, |
| { |
| "key": "mmlu_pro", |
| "hf_repo": "TIGER-Lab/MMLU-Pro", |
| "revision": "b189ec765aa7ed75c8acfea42df31fdae71f97be", |
| "files": ["data/test-00000-of-00001.parquet"], |
| "format": "parquet", |
| "license": "mit", |
| "redistributable": true, |
| "exposure": "likely", |
| "tier": "mcq_wide", |
| "task_type": "choice", |
| "n_options": 10, |
| "columns": ["question_id", "question", "options", "answer", "answer_index", "cot_content", "category", "src"], |
| "model_visible": ["question", "options"], |
| "gold_field": "answer_index", |
| "gold_vocab": null, |
| "hidden": ["answer", "answer_index", "cot_content"], |
| "stratify_by": "category", |
| "target_n": 500, |
| "notes": "Option count varies per row and is up to 10. Record the actual per-row count as n_options; do not assume 10. cot_content is a leaked chain of thought and must never enter state." |
| }, |
| { |
| "key": "banking77", |
| "hf_repo": "mteb/banking77", |
| "revision": "18072d2685ea682290f7b8924d94c62acc19c0b2", |
| "files": ["data/test-00000-of-00001.parquet"], |
| "format": "parquet", |
| "license": "mit", |
| "redistributable": true, |
| "exposure": "reported", |
| "tier": "options_stress", |
| "task_type": "choice", |
| "n_options": 77, |
| "columns": ["text", "label", "label_text"], |
| "model_visible": ["text"], |
| "gold_field": "label_text", |
| "gold_vocab": null, |
| "hidden": ["label", "label_text"], |
| "stratify_by": "label_text", |
| "target_n": 500, |
| "expected_n_classes": 77, |
| "notes": "SUBSTITUTION: PolyAI/banking77 and legacy-datasets/banking77 are loading-script datasets with no parquet, which current datasets versions will not load without remote code execution. mteb/banking77 carries the same test split as parquet plus label_text, so the 77 intent names are read from the data instead of being typed from memory. Upstream data is CC-BY-4.0 from PolyAI; cite PolyAI, not mteb, in the paper. Together's published recipe reports BANKING77 in its training mixture, hence exposure=reported." |
| }, |
| { |
| "key": "pubmedqa", |
| "hf_repo": "qiaojin/PubMedQA", |
| "revision": "9001f2853fb87cab8d220904e0de81ac6973b318", |
| "files": ["pqa_labeled/train-00000-of-00001.parquet"], |
| "format": "parquet", |
| "license": "mit", |
| "redistributable": true, |
| "exposure": "unknown", |
| "tier": "grounded", |
| "task_type": "choice", |
| "n_options": 3, |
| "columns": ["pubid", "question", "context", "long_answer", "final_decision"], |
| "model_visible": ["question", "context"], |
| "gold_field": "final_decision", |
| "gold_vocab": ["yes", "no", "maybe"], |
| "hidden": ["final_decision", "long_answer"], |
| "stratify_by": "final_decision", |
| "target_n": 500, |
| "notes": "pqa_labeled is the expert-annotated subset and the only one with trustworthy gold. Its split is named train upstream even though it is the evaluation set. context is a struct holding contexts/labels/meshes; pass only contexts. long_answer is the gold rationale and stays hidden." |
| }, |
| { |
| "key": "sst5", |
| "hf_repo": "SetFit/sst5", |
| "revision": "e51bdcd8cd3a30da231967c1a249ba59361279a3", |
| "files": ["test.jsonl"], |
| "format": "jsonl", |
| "license": "unspecified-on-card", |
| "redistributable": false, |
| "exposure": "reported", |
| "tier": "ordinal", |
| "task_type": "score", |
| "n_options": 5, |
| "columns": ["text", "label", "label_text"], |
| "model_visible": ["text"], |
| "gold_field": "label", |
| "gold_vocab": ["very negative", "negative", "neutral", "positive", "very positive"], |
| "hidden": ["label", "label_text"], |
| "stratify_by": "label_text", |
| "target_n": 500, |
| "notes": "Ordered 0-4, verified from the data, so the score rubric order is measured not assumed. The card states no license, so ship case ids plus row indices and a content hash rather than the sentence text in a permissive release. Upstream SST is from Stanford; check its terms before redistribution. Together's recipe reports SST-5, hence exposure=reported." |
| }, |
| { |
| "key": "atbench500", |
| "hf_repo": "AI45Research/ATBench", |
| "revision": "4476ef92ed8f85c8d58d8a5b9dfdf55aa7893138", |
| "files": ["ATBench500/test.json"], |
| "format": "json_array", |
| "license": "apache-2.0", |
| "redistributable": true, |
| "exposure": "unknown", |
| "tier": "safety", |
| "task_type": "noul", |
| "n_options": 2, |
| "columns": ["tool_used", "content", "conv_id", "label", "risk_source", "failure_mode", "real_world_harm"], |
| "model_visible": ["tool_used", "content"], |
| "gold_field": "label", |
| "gold_vocab": [0, 1], |
| "hidden": ["label", "risk_source", "failure_mode", "real_world_harm"], |
| "stratify_by": "label", |
| "target_n": 500, |
| "notes": "label 0/1 for safe/unsafe: confirm the polarity against the dataset card before scoring, the direction is not self-evident from the values. risk_source, failure_mode and real_world_harm are gold-adjacent annotations and are the classic accidental leak in this dataset. content is a nested list of trajectory turns and is long: the card reports about 1.52k tokens average, which will collide with a 512-token context." |
| }, |
| { |
| "key": "aegis2", |
| "hf_repo": "nvidia/Aegis-AI-Content-Safety-Dataset-2.0", |
| "revision": "d86bb8bedff51d25ac834ab7838f1cc61acb7a2c", |
| "files": ["test.json"], |
| "format": "json_lines_or_array", |
| "license": "cc-by-4.0", |
| "redistributable": true, |
| "exposure": "unknown", |
| "tier": "safety", |
| "task_type": "noul", |
| "n_options": 2, |
| "columns": ["id", "prompt", "response", "prompt_label", "response_label", "violated_categories", "prompt_label_source", "response_label_source"], |
| "model_visible": ["prompt", "response"], |
| "gold_field": "prompt_label", |
| "gold_vocab": ["safe", "unsafe"], |
| "hidden": ["prompt_label", "response_label", "violated_categories", "prompt_label_source", "response_label_source"], |
| "stratify_by": "prompt_label", |
| "target_n": 500, |
| "notes": "response and response_label are null on many rows, so run prompt-level and response-level scoring as two separate slices rather than one blended number. violated_categories doubles as a choice-type taxonomy task later." |
| }, |
| { |
| "key": "prompt_injections", |
| "hf_repo": "deepset/prompt-injections", |
| "revision": "4f61ecb038e9c3fb77e21034b22511b523772cdd", |
| "files": ["data/test-00000-of-00001-701d16158af87368.parquet"], |
| "format": "parquet", |
| "license": "apache-2.0", |
| "redistributable": true, |
| "exposure": "unknown", |
| "tier": "safety", |
| "task_type": "noul", |
| "n_options": 2, |
| "columns": ["text", "label"], |
| "model_visible": ["text"], |
| "gold_field": "label", |
| "gold_vocab": [0, 1], |
| "hidden": ["label"], |
| "stratify_by": "label", |
| "target_n": null, |
| "notes": "Small enough to use whole, roughly 11 kB of test parquet. Multilingual and small, so treat it as a diagnostic, not an injection-defence certification. Content is adversarial by construction: it is data, never instructions, and the adapter must wrap it as untrusted." |
| }, |
| { |
| "key": "jailbreak_classification", |
| "hf_repo": "jackhhao/jailbreak-classification", |
| "revision": "2f2ceeb39658696fd3f462403562b6eea5306287", |
| "files": ["default/jailbreak_dataset_test.csv"], |
| "format": "csv", |
| "license": "apache-2.0", |
| "redistributable": true, |
| "exposure": "unknown", |
| "tier": "safety", |
| "task_type": "noul", |
| "n_options": 2, |
| "columns": ["prompt", "type"], |
| "model_visible": ["prompt"], |
| "gold_field": "type", |
| "gold_vocab": ["benign", "jailbreak"], |
| "hidden": ["type"], |
| "stratify_by": "type", |
| "target_n": 500, |
| "notes": "A balanced/ variant also exists upstream; default/ is pinned here. Same untrusted-data handling as prompt_injections." |
| }, |
| { |
| "key": "vqa_rad", |
| "hf_repo": "flaviagiammarino/vqa-rad", |
| "revision": "bcf91e7654fb9d51c8ab6a5b82cacf3fafd2fae9", |
| "files": ["data/test-00000-of-00001-e5bc3d208bb4deeb.parquet"], |
| "format": "parquet", |
| "license": "cc0-1.0", |
| "redistributable": true, |
| "exposure": "unknown", |
| "tier": "multimodal", |
| "task_type": "noul", |
| "n_options": 2, |
| "columns": ["image", "question", "answer"], |
| "model_visible": ["image", "question"], |
| "gold_field": "answer", |
| "gold_vocab": null, |
| "hidden": ["answer"], |
| "stratify_by": null, |
| "target_n": null, |
| "notes": "Answers are free text. Filter to the closed-form yes/no subset for a clean noul slice and report how many rows were dropped. Images are embedded in the parquet. CC0 means the adapted slice is republishable, which is rare for medical imaging. djev only: the Jev API path in this workspace has no image input wired." |
| }, |
| { |
| "key": "scienceqa", |
| "hf_repo": "derek-thomas/ScienceQA", |
| "revision": "f18b0a70359ebfb41f658fd564208d0355b013f4", |
| "files": ["data/test-00000-of-00001-f0e719df791966ff.parquet"], |
| "format": "parquet", |
| "license": "cc-by-sa-4.0", |
| "redistributable": true, |
| "exposure": "likely", |
| "tier": "multimodal", |
| "task_type": "choice", |
| "n_options": null, |
| "columns": ["image", "question", "choices", "answer", "hint", "task", "grade", "subject", "topic", "category", "skill", "lecture", "solution"], |
| "model_visible": ["image", "question", "choices", "hint"], |
| "gold_field": "answer", |
| "gold_vocab": null, |
| "hidden": ["answer", "lecture", "solution"], |
| "stratify_by": "subject", |
| "target_n": 500, |
| "notes": "Largest download at roughly 122 MB for the test parquet. image is null on many rows, which gives a free text-only versus image-present ablation on identical adapter code. lecture and solution are gold rationales and must stay hidden. answer is an index into choices, so option count varies per row." |
| } |
| ], |
| "excluded": [ |
| {"key": "toxic_chat", "hf_repo": "lmsys/toxic-chat", "reason": "cc-by-nc-4.0. Research use only, cannot sit inside a permissively licensed release."}, |
| {"key": "wildguardmix", "hf_repo": "allenai/wildguardmix", "reason": "Gated (auto). Requires accepting terms on the HF account first; add later if wanted."}, |
| {"key": "banking77_polyai", "hf_repo": "PolyAI/banking77", "reason": "Loading-script dataset with no parquet files. Superseded by mteb/banking77 carrying the same upstream data."} |
| ], |
| "local_tier0": [ |
| {"key": "jevbench_easy", "path": "../quick_benchmark/data/jevbench_easy_48.jsonl", "use": "adapter sanity check against published Jev numbers. Not a results table."}, |
| {"key": "rjudge50", "path": "../quick_benchmark/data/inputs.json", "use": "offline safety smoke test, gold in answers.json."} |
| ] |
| } |
|
|
