Publish reusable historical exclusion evidence

README.md

@@ -308,3 +308,8 @@ TypeSafe and OpenJev.
 ## Incremental filtered WebInstruct addition
 
 Filtered `tasksource/webinstruct` configs `mc` and `binary`, pinned at `05e8b528d68c628c1af682d47bdb53f85535876a`, are appended as standalone shards to both `default` and `full`. The original train/test split assignment is preserved; validation is unchanged. All choices of included rows are retained, with canonical Jev option permutation and one-hot targets; 47 MC rows with duplicate option text are omitted under the existing choice contract. binary questions use explicit no/false and yes/true choices. No repairs or derived variants are included. Existing shards and mixture sampling are unchanged. See `webinstruct-addition.json` for exact counts, transformations and provenance.
+
+
+## Reusable selection evidence
+
+The [exclusion evidence resource](quality/filtered-full-v1/README.md) records historical filtered-full selection membership separately from licensing and benchmark policies. Remaining candidates are unattributed exclusions, not verified errors. Canonical decision keys support evidence reuse across recasts.

quality/filtered-full-v1/README.md

@@ -0,0 +1,84 @@
+# Reusable exclusion evidence
+
+This resource records selection evidence from the historical `filtered-full`
+configuration. It is independent of the ongoing DeepSeek audit. No new model
+calls, relabeling, approved removal list or training-data change is involved.
+
+Compare `full/train` at `d2ab1d12be4463fb7ac1f877ed2193b3bb59642a` with
+`filtered-full/train` at `84008c8a8f4640a2b8a3b222cb62f23ae5fa6e73` by their
+original unique decision `id`. Frozen source files are checksum-verified. Every
+selected ID must exist in the source and selection counts must reconcile.
+
+| Observed exclusion category | Rows |
+| --- | ---: |
+| Non-commercial or unspecified license policy | 1,025,115 |
+| Explicit benchmark-source holdout policy | 122,480 |
+| Unattributed remaining exclusions | 46,927 |
+| All excluded decisions | 1,194,522 |
+
+License and benchmark attribution reproduces the published rules and reconciles
+their exact counts. These are training policies, not source-label defects.
+The policy categories are mutually exclusive: benchmark holdout takes precedence
+over licensing. Both rules apply to 68,655 rows; this precedence reproduces the
+report's attribution without counting those rows twice.
+The remaining 46,927 rows combine 23,138 reported GLM rejections, 7 unresolved
+API errors and 23,782 overlap, positional-option, budget and conflicting-target
+exclusions. The published release supplies aggregate counts, not the per-row
+GLM or other remaining verdicts. Their individual attribution cannot be inferred
+uniquely from selection membership. None is labeled a verified error here.
+
+## Files
+
+- `excluded.parquet`: every excluded decision's original ID, source, group,
+  question ID, variant, split, license-use policy, reconstructed category and
+  candidate decision key where available.
+- `candidates.parquet`: the 46,927 unattributed exclusions, with complete original
+  rows and canonical decision keys. This is a review resource, not a blacklist.
+- `manifest.json`: pinned revisions, fingerprints, category counts, per-source
+  counts, reconstructed benchmark rules and explicit attribution limitations.
+
+The Hub directory is `quality/filtered-full-v1/`. These are sidecar resources;
+existing dataset configurations and Parquet shards remain unchanged.
+
+## Reuse across recasts
+
+`tasksource.quality.decision_key` accepts published typed decisions and canonical
+Tasksource `recast='jev'` rows. The SHA256 identity includes the task, exact input,
+question, decision kind and semantic gold/distribution. Neutral choice order is
+canonicalized; score order and positional option references remain significant.
+Corrected gold or context, different questions, and different label vocabularies
+do not silently inherit an earlier flag. IDs based on row position are insufficient
+when sampling or preprocessing changes.
+
+```python
+from huggingface_hub import hf_hub_download
+from tasksource.quality import ExclusionIndex, decision_key
+
+path = hf_hub_download(
+    'tasksource/tasksource-jev-typed-decisions',
+    'quality/filtered-full-v1/candidates.parquet', repo_type='dataset',
+)
+evidence = ExclusionIndex.from_parquet(path)
+matches = evidence.match(canonical_decision)
+# Matches indicate an unattributed historical exclusion, not an approved defect.
+```
+
+For another instruct/chat rendering, attach `decision_key(canonical_decision)`
+before formatting and carry that identity with the unchanged example; query it
+with `evidence.lookup(key)`. Recompute it if context, task or gold changes. If an
+old recast did not preserve that identity and changes question wording or input
+formatting, an explicit adapter or source mapping is needed. There is no fuzzy
+matching or automatic group-wide propagation. A rejected derived question does
+not establish that every question in its group is bad.
+
+## Reproduction
+
+```bash
+PYTHONPATH=.:src python scripts/extract_jev_exclusions.py
+```
+
+The script exports exact set membership, reproduces policy categories and validates
+source hashes and selection coverage. Per-row reviewer reasons and adjudication
+can later be attached to these identities without recasting all training data.
+The ongoing audit's model flags still require review before becoming an approved
+cross-recast removal resource.

quality/filtered-full-v1/candidates.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b29c296b6e7e7d78d6ccdf9a50a65e76e864ea511e8affd73886742d1d625c3
+size 53743730

quality/filtered-full-v1/excluded.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8edd2abcf71352682d2d5b1a7eda0f67136e3d644060290630155055fe1d8374
+size 41681395

quality/filtered-full-v1/manifest.json

@@ -0,0 +1,685 @@
+{
+  "schema_version": 1,
+  "repository": "tasksource/tasksource-jev-typed-decisions",
+  "release_revision": "84008c8a8f4640a2b8a3b222cb62f23ae5fa6e73",
+  "source_revision": "d2ab1d12be4463fb7ac1f877ed2193b3bb59642a",
+  "source_config": "full",
+  "source_split": "train",
+  "selected_config": "filtered-full",
+  "source_rows": 2500000,
+  "retained_rows": 1305478,
+  "excluded_rows": 1194522,
+  "categories": {
+    "license_policy": 1025115,
+    "benchmark_policy": 122480,
+    "unattributed_exclusion": 46927
+  },
+  "candidate_rows": 46927,
+  "llm_rejected_rows_reported": 23138,
+  "unresolved_api_errors_reported": 7,
+  "decision_key": "tasksource-decision-v1; SHA256; exact task/input/question/gold; neutral choice order canonicalized",
+  "identified_error_ids_available": false,
+  "approved_removals": [],
+  "attribution": "Set difference proves exclusion membership. License/benchmark policy predicates are reproduced from the report. Remaining exclusions cannot be assigned individual GLM/overlap/budget/conflict reasons.",
+  "matching": "Decision-level content lookup only; no group-wide, row-index-only or fuzzy transfer. Carry the canonical key before changing recast rendering. Corrected gold/context no longer matches.",
+  "benchmark_rule": {
+    "blocked_roots": [
+      "acos",
+      "ai2_arc",
+      "anli",
+      "api-bank",
+      "arc-challenge",
+      "arc-easy",
+      "banking77",
+      "bbh",
+      "belebele",
+      "bfcl",
+      "bigbench",
+      "blimp",
+      "bright",
+      "chessbench",
+      "circa",
+      "cladder",
+      "clinc150",
+      "clinc_oos",
+      "commonsense_qa",
+      "contract-nli",
+      "contractnli",
+      "cos_e",
+      "cruxeval",
+      "esci",
+      "financial_phrasebank",
+      "finentity",
+      "forecastbench",
+      "gpqa",
+      "gsm8k",
+      "hellaswag",
+      "hle",
+      "hover",
+      "hover-3way",
+      "humicroedit",
+      "isarcasmeval",
+      "it-support-tickets",
+      "jevbench",
+      "judgebench",
+      "mmlu",
+      "mmlu-pro",
+      "musr",
+      "nli4ct_semeval2024",
+      "openbookqa",
+      "phishnchips",
+      "ragtruth",
+      "routerbench",
+      "sata-bench",
+      "schema_guided_dstc8",
+      "sgd",
+      "sgd-x",
+      "simplebench",
+      "tabfact",
+      "toolret",
+      "truthfulqa",
+      "vast",
+      "when2call",
+      "winogrande"
+    ],
+    "root_matching": "Take the part before the first slash, casefold, then remove non-alphanumeric ASCII characters before comparing.",
+    "casefolded_source_prefixes": [
+      "multilingual/xcsr/x-csqa-"
+    ],
+    "implementation_sha256": "fa3ed2826c1d9b0a40f6cc4102ebd46c2609ba8100b0d1d80be9293764cabe57"
+  },
+  "policy_precedence": [
+    "benchmark_policy",
+    "license_policy"
+  ],
+  "per_source_excluded": {
+    "AES2-essay-scoring": 3080,
+    "AdjectiveScaleProbe-nli": 5017,
+    "AmbigNQ-clarifying-question": 3965,
+    "BeaverTails": 4331,
+    "CONDAQA": 483,
+    "CREAK": 101,
+    "ConTRoL-nli": 8864,
+    "Dilemmas_Disagreement": 2288,
+    "Dynasent_Disagreement": 2415,
+    "FLD.v2/default": 2955,
+    "FLD.v2/star": 2918,
+    "FLUTE": 31,
+    "FOL-nli": 1139,
+    "HatemojiBuild": 410,
+    "HelpSteer2/complexity": 2,
+    "HelpSteer2/correctness": 5,
+    "HelpSteer2/helpfulness": 1,
+    "HelpSteer2/verbosity": 1,
+    "HelpSteer3/edit_quality": 7,
+    "HelpSteer3/feedback": 3,
+    "HelpSteer3/preference": 11,
+    "HelpSteer3/preference_strength": 3,
+    "HelpSteer3/principle": 1227,
+    "I2D2": 197,
+    "IntentGrasp/all": 15978,
+    "LogicNLI": 5039,
+    "MOH": 1749,
+    "MSciNLI": 318,
+    "MedQA-USMLE-4-options-hf": 210,
+    "PARADISE": 150,
+    "PARARULE-Plus": 3,
+    "PKU-SafeRLHF/helpfulness": 8671,
+    "PKU-SafeRLHF/safety": 8710,
+    "Pol_NLI": 4340,
+    "Politeness_Disagreement": 1574,
+    "Prompt-injection-dataset/full": 91,
+    "PromptShield": 206,
+    "ReSQ": 2535,
+    "SBIC_Disagreement": 2365,
+    "SChem_Disagreement": 2000,
+    "SDOH-NLI": 11,
+    "SHP": 8730,
+    "Sarcasm_News_Headline": 4266,
+    "ScienceQA_text_only": 7483,
+    "ShellRisk-Bench": 4050,
+    "SpaRTUN": 12309,
+    "SpaceNLI": 133,
+    "Touche23-ValueEval": 12,
+    "TuringBench": 5,
+    "UNLI": 33,
+    "UltraFeedback-paired": 128,
+    "VUAC": 2186,
+    "WANLI": 571,
+    "acceptability-prediction/binary_votes": 94,
+    "acceptability-prediction/rating_votes": 271,
+    "add_one_rte": 3877,
+    "ade_corpus_v2/Ade_corpus_v2_classification": 4119,
+    "ag_news": 5285,
+    "agent_action_safety": 16,
+    "ai2_arc/ARC-Challenge/challenge": 1303,
+    "ai2_arc/ARC-Easy/challenge": 2600,
+    "amazon_counterfactual/en": 160,
+    "ambient": 61,
+    "anli/a1": 5193,
+    "anli/a2": 5212,
+    "anli/a3": 5202,
+    "app_reviews": 5134,
+    "apt": 3869,
+    "arct": 61,
+    "args_me": 155,
+    "argument-feedback": 2880,
+    "art": 12177,
+    "auditor_review": 3885,
+    "autotnli": 170,
+    "avicenna": 3880,
+    "babi_nli/basic-induction": 2,
+    "babi_nli/counting": 1,
+    "babi_nli/lists-sets": 1,
+    "babi_nli/path-finding": 3,
+    "babi_nli/positional-reasoning": 8,
+    "babi_nli/size-reasoning": 1,
+    "babi_nli/three-supporting-facts": 9,
+    "babi_nli/two-arg-relations": 7,
+    "balanced-copa": 22,
+    "banking77": 6770,
+    "biosift-nli": 4361,
+    "blog_authorship_corpus/age": 13,
+    "blog_authorship_corpus/gender": 16,
+    "blog_authorship_corpus/job": 50,
+    "boolq-natural-perturbations": 3982,
+    "brainteasers/SP": 410,
+    "brainteasers/WP": 412,
+    "breaking_nli": 144,
+    "chaos-mnli-ambiguity/votes": 1439,
+    "chatbot_arena_conversations": 10158,
+    "chemprot/chemprot_full_source": 5774,
+    "cicero": 1025,
+    "circa": 5325,
+    "citation_intent": 1632,
+    "civil_comments/identity_attack_share": 12,
+    "civil_comments/insult_share": 111,
+    "civil_comments/obscene_share": 8,
+    "civil_comments/severe_toxicity_share": 1,
+    "civil_comments/sexual_explicit_share": 4,
+    "civil_comments/threat_share": 12,
+    "civil_comments/toxicity_share": 10,
+    "cladder": 3942,
+    "clcd-english": 186,
+    "clinc_oos/plus": 6861,
+    "cloth": 212,
+    "clutrr": 5907,
+    "cnli": 4890,
+    "codah/codah": 26,
+    "code_x_glue_cc_defect_detection": 4183,
+    "com2sense": 2044,
+    "commonsense_qa": 10401,
+    "commonsense_qa_2.0": 63,
+    "conceptrules_v2": 64,
+    "conj_nli": 5011,
+    "conll2003/ner_tags": 30742,
+    "contract-nli/contractnli_a/seg": 2604,
+    "contract-nli/contractnli_b/full": 2377,
+    "corr2cause": 2671,
+    "cos_e/v1.0": 8727,
+    "cosmos_qa": 249,
+    "counterfactually-augmented-imdb": 2115,
+    "counterfactually-augmented-snli": 4879,
+    "crowdflower/airline-sentiment": 728,
+    "crowdflower/corporate-messaging": 802,
+    "crowdflower/political-media-audience": 784,
+    "crowdflower/political-media-bias": 792,
+    "crowdflower/political-media-message": 827,
+    "crowdflower/sentiment_nuclear_power": 215,
+    "crowdflower/text_emotion": 819,
+    "crowdflower/tweet_global_warming": 811,
+    "cycic_classification": 39,
+    "cycic_multiplechoice": 202,
+    "dadc-limit-nli": 898,
+    "dataset_train_nli": 19432,
+    "dbpedia_14/dbpedia_14": 19,
+    "defeasible-nli/atomic": 23,
+    "defeasible-nli/snli": 9,
+    "defeasible-nli/social": 10,
+    "definite_pronoun_resolution": 1371,
+    "dgen": 25,
+    "discosense": 9653,
+    "discovery/discovery": 118,
+    "disrpt/eng.dep.scidtb.rels": 32,
+    "dnc": 4360,
+    "doc-nli": 618,
+    "docred": 154,
+    "dream": 5462,
+    "dynasent/r1_votes": 187,
+    "dynasent/r2_votes": 41,
+    "e-CARE": 334,
+    "ekar_english": 947,
+    "emo/emo2019": 15,
+    "emotion": 5974,
+    "english-grading/cohesion": 444,
+    "english-grading/conventions": 459,
+    "english-grading/grammar": 449,
+    "english-grading/phraseology": 453,
+    "english-grading/syntax": 436,
+    "english-grading/vocabulary": 451,
+    "equate": 78,
+    "esci": 2976,
+    "ethics/commonsense": 6,
+    "ethics/deontology": 1,
+    "ethics/justice": 9,
+    "ethics/virtue": 5,
+    "ethos/binary": 1161,
+    "ethos/multilabel": 2097,
+    "feasibilityQA": 169,
+    "fever-evidence-related": 1,
+    "few_rel/default": 4367,
+    "fig-qa": 146,
+    "financial_phrasebank/sentences_allagree": 2274,
+    "folio": 145,
+    "fool-me-twice": 93,
+    "fracas": 315,
+    "gen_debiased_nli/mnli_par_z": 35,
+    "gen_debiased_nli/mnli_seq_z": 30,
+    "gen_debiased_nli/mnli_z_aug": 35,
+    "gen_debiased_nli/snli_par_z": 14,
+    "gen_debiased_nli/snli_seq_z": 20,
+    "gen_debiased_nli/snli_z_aug": 22,
+    "github-issue-similarity": 147,
+    "glue/cola": 774,
+    "glue/mnli": 778,
+    "glue/mrpc": 802,
+    "glue/qnli": 786,
+    "glue/qqp": 823,
+    "glue/rte": 767,
+    "glue/sst2": 753,
+    "glue/stsb": 5706,
+    "glue/wnli": 743,
+    "go_emotions/simplified": 2,
+    "goal-step-wikihow/goal": 143,
+    "goal-step-wikihow/order": 102,
+    "goal-step-wikihow/step": 229,
+    "google_wellformed_query": 4,
+    "hans": 4336,
+    "hate_speech18": 5,
+    "hate_speech_offensive": 5146,
+    "head_qa/en": 80,
+    "headline_cause/en_simple": 59,
+    "health_fact": 118,
+    "hellaswag": 14239,
+    "help-nli": 746,
+    "hh-rlhf/harmless-base": 34,
+    "hh-rlhf/helpful-base": 7,
+    "hh-rlhf/helpful-rejection-sampled": 7,
+    "hlgd": 1031,
+    "hope_edi/english": 100,
+    "hover": 4035,
+    "humicroedit/subtask-1": 1265,
+    "humicroedit/subtask-2": 1407,
+    "hyperpartisan_news": 58,
+    "idioms-nli": 4138,
+    "imdb": 4184,
+    "implicatures": 29,
+    "implicit-hate-stg1": 5069,
+    "insincere-questions": 4350,
+    "it-support-tickets": 1826,
+    "jigsaw_toxicity": 5,
+    "joci": 5178,
+    "language-identification": 3344,
+    "lewidi/hs_brexit": 784,
+    "lewidi/md_agreement": 5376,
+    "lewidi/mp": 5374,
+    "lex_glue/case_hold": 321,
+    "lex_glue/ledgar": 39,
+    "lex_glue/scotus": 1044,
+    "lex_glue/unfair_tos": 147,
+    "lexical_relation_classification/BLESS": 1176,
+    "lexical_relation_classification/CogALexV": 1155,
+    "lexical_relation_classification/EVALution": 1185,
+    "lexical_relation_classification/K&H+N": 1190,
+    "lexical_relation_classification/ROOT09": 1207,
+    "liar": 2996,
+    "lifecycle-entailment": 3936,
+    "lingnli": 5254,
+    "linguisticprobing/bigram_shift": 533,
+    "linguisticprobing/coordination_inversion": 601,
+    "linguisticprobing/obj_number": 546,
+    "linguisticprobing/odd_man_out": 572,
+    "linguisticprobing/past_present": 564,
+    "linguisticprobing/subj_number": 579,
+    "linguisticprobing/top_constituents": 544,
+    "logical-entailment": 69,
+    "logical-fallacy": 2903,
+    "logiqa": 7591,
+    "logiqa-2.0-nli": 5800,
+    "lonli": 202,
+    "lsat-ar": 1821,
+    "lsat-rc": 2079,
+    "lsat_qa/all": 236,
+    "math_qa": 2143,
+    "mc_taco": 4329,
+    "mctest-nli": 8127,
+    "medical_questions_pairs": 3521,
+    "medmcqa": 602,
+    "mindgames": 723,
+    "missing-item-prediction/contrastive": 1991,
+    "monli": 1162,
+    "monotonicity-entailment": 174,
+    "moral_stories/full": 47,
+    "mpe": 4895,
+    "multilingual/MLMA_hate_speech": 4,
+    "multilingual/NusaX-senti/bug": 1,
+    "multilingual/NusaX-senti/ind": 5,
+    "multilingual/amazon_reviews_multi/all_languages": 5141,
+    "multilingual/americas_nli/all_languages": 63,
+    "multilingual/clue/afqmc": 2061,
+    "multilingual/clue/ocnli": 2164,
+    "multilingual/clue/tnews": 2025,
+    "multilingual/disrpt/deu.rst.pcc.rels": 11,
+    "multilingual/disrpt/eus.rst.ert.rels": 8,
+    "multilingual/disrpt/fas.rst.prstc.rels": 18,
+    "multilingual/disrpt/fra.sdrt.annodis.rels": 6,
+    "multilingual/disrpt/nld.rst.nldt.rels": 4,
+    "multilingual/disrpt/por.rst.cstn.rels": 5,
+    "multilingual/disrpt/rus.rst.rrt.rels": 6,
+    "multilingual/disrpt/spa.rst.rststb.rels": 3,
+    "multilingual/disrpt/tha.pdtb.tdtb.rels": 2,
+    "multilingual/disrpt/zho.rst.gcdt.rels": 5,
+    "multilingual/exams/multilingual": 475,
+    "multilingual/indic_glue/actsa-sc.te/sentiment": 839,
+    "multilingual/indic_glue/bbca.hi/news": 830,
+    "multilingual/indic_glue/iitp-mr.hi/sentiment": 798,
+    "multilingual/indic_glue/iitp-pr.hi/sentiment": 793,
+    "multilingual/indic_glue/inltkh": 889,
+    "multilingual/indic_glue/inltkh.te/sentiment": 767,
+    "multilingual/indic_glue/md.hi/discourse_mode": 880,
+    "multilingual/indic_glue/sna.bn/news": 901,
+    "multilingual/indic_glue/wstp": 13802,
+    "multilingual/klue/nli": 13,
+    "multilingual/klue/sts": 32,
+    "multilingual/klue/ynat": 41,
+    "multilingual/language-identification": 3329,
+    "multilingual/massive": 166,
+    "multilingual/miam": 116,
+    "multilingual/mms": 5158,
+    "multilingual/mtop": 107,
+    "multilingual/multilingual-NLI-26lang-2mil7": 5283,
+    "multilingual/multilingual-sentiments/all": 6,
+    "multilingual/oasst1_pairwise_rlhf_reward": 10031,
+    "multilingual/offenseval_2020/ar": 1144,
+    "multilingual/offenseval_2020/da": 1154,
+    "multilingual/offenseval_2020/gr": 1156,
+    "multilingual/offenseval_2020/tr": 1137,
+    "multilingual/offenseval_dravidian/kannada": 47,
+    "multilingual/offenseval_dravidian/malayalam": 18,
+    "multilingual/offenseval_dravidian/tamil": 41,
+    "multilingual/paws-x/de": 1085,
+    "multilingual/paws-x/en": 1070,
+    "multilingual/paws-x/es": 1093,
+    "multilingual/paws-x/fr": 1100,
+    "multilingual/paws-x/ja": 1093,
+    "multilingual/paws-x/ko": 1069,
+    "multilingual/paws-x/zh": 1059,
+    "multilingual/tweet_sentiment_multilingual": 8,
+    "multilingual/tydi-as2-balanced": 4343,
+    "multilingual/universal-joy": 5,
+    "multilingual/wili_2018": 7055,
+    "multilingual/x-fact": 3461,
+    "multilingual/x-stance": 1195,
+    "multilingual/xcopa/et": 1,
+    "multilingual/xcopa/id": 1,
+    "multilingual/xcopa/it": 1,
+    "multilingual/xcopa/qu": 1,
+    "multilingual/xcopa/sw": 1,
+    "multilingual/xcopa/th": 2,
+    "multilingual/xcopa/tr": 4,
+    "multilingual/xcopa/translation-et": 1,
+    "multilingual/xcopa/translation-ht": 1,
+    "multilingual/xcopa/translation-id": 2,
+    "multilingual/xcopa/translation-ta": 3,
+    "multilingual/xcopa/translation-th": 2,
+    "multilingual/xcopa/translation-tr": 4,
+    "multilingual/xcsr/X-CODAH-ar": 5,
+    "multilingual/xcsr/X-CODAH-de": 3,
+    "multilingual/xcsr/X-CODAH-en": 1,
+    "multilingual/xcsr/X-CODAH-es": 1,
+    "multilingual/xcsr/X-CODAH-fr": 2,
+    "multilingual/xcsr/X-CODAH-hi": 3,
+    "multilingual/xcsr/X-CODAH-it": 2,
+    "multilingual/xcsr/X-CODAH-jap": 2,
+    "multilingual/xcsr/X-CODAH-pl": 7,
+    "multilingual/xcsr/X-CODAH-pt": 3,
+    "multilingual/xcsr/X-CODAH-ru": 4,
+    "multilingual/xcsr/X-CODAH-sw": 4,
+    "multilingual/xcsr/X-CODAH-ur": 2,
+    "multilingual/xcsr/X-CODAH-vi": 4,
+    "multilingual/xcsr/X-CODAH-zh": 3,
+    "multilingual/xcsr/X-CSQA-ar": 488,
+    "multilingual/xcsr/X-CSQA-de": 481,
+    "multilingual/xcsr/X-CSQA-en": 481,
+    "multilingual/xcsr/X-CSQA-es": 486,
+    "multilingual/xcsr/X-CSQA-fr": 495,
+    "multilingual/xcsr/X-CSQA-hi": 482,
+    "multilingual/xcsr/X-CSQA-it": 491,
+    "multilingual/xcsr/X-CSQA-jap": 489,
+    "multilingual/xcsr/X-CSQA-nl": 482,
+    "multilingual/xcsr/X-CSQA-pl": 485,
+    "multilingual/xcsr/X-CSQA-pt": 477,
+    "multilingual/xcsr/X-CSQA-ru": 488,
+    "multilingual/xcsr/X-CSQA-sw": 494,
+    "multilingual/xcsr/X-CSQA-ur": 496,
+    "multilingual/xcsr/X-CSQA-vi": 480,
+    "multilingual/xcsr/X-CSQA-zh": 491,
+    "multilingual/xglue/nc": 1491,
+    "multilingual/xglue/qadsm": 1466,
+    "multilingual/xglue/qam": 1474,
+    "multilingual/xglue/wpr": 1341,
+    "multilingual/xlwic/xlwic_de_de": 1222,
+    "multilingual/xlwic/xlwic_en_ko": 1268,
+    "multilingual/xlwic/xlwic_fr_fr": 1235,
+    "multilingual/xlwic/xlwic_it_it": 1244,
+    "multilingual/xnli": 14721,
+    "multilingual/xstory_cloze/id": 1,
+    "multilingual/xstory_cloze/my": 2,
+    "multilingual/xstory_cloze/te": 3,
+    "multilingual/xstory_cloze/zh": 3,
+    "mutual": 7353,
+    "nan-nli": 89,
+    "natural-language-satisfiability": 3936,
+    "naturallogic": 482,
+    "nlgraph": 3589,
+    "nli-veridicality-transitivity": 4326,
+    "nli4ct_semeval2024": 2308,
+    "nli_fever": 5268,
+    "nope": 2737,
+    "numer_sense": 113,
+    "oasst2/creativity": 8980,
+    "oasst2/helpfulness": 8980,
+    "oasst2/humor": 8980,
+    "oasst2/quality": 8979,
+    "oasst2/toxicity": 8979,
+    "oasst2/violence": 8979,
+    "oasst2_pairwise_rlhf_reward": 11595,
+    "onestop_qa": 62,
+    "open_question_type": 87,
+    "openbookqa": 5256,
+    "parade": 3948,
+    "patent-phrase-similarity": 231,
+    "path-naturalness-prediction": 12083,
+    "paws/labeled_final": 2443,
+    "paws/labeled_swap": 2413,
+    "persuasion": 1390,
+    "phrase_similarity": 3931,
+    "piqa": 9914,
+    "pragmeval/emergent": 408,
+    "pragmeval/emobank-arousal": 381,
+    "pragmeval/emobank-dominance": 357,
+    "pragmeval/emobank-valence": 380,
+    "pragmeval/gum": 355,
+    "pragmeval/mrda": 356,
+    "pragmeval/pdtb": 415,
+    "pragmeval/persuasiveness-claimtype": 210,
+    "pragmeval/persuasiveness-eloquence": 394,
+    "pragmeval/persuasiveness-premisetype": 385,
+    "pragmeval/persuasiveness-relevance": 383,
+    "pragmeval/persuasiveness-specificity": 393,
+    "pragmeval/persuasiveness-strength": 404,
+    "pragmeval/sarcasm": 414,
+    "pragmeval/squinky-formality": 383,
+    "pragmeval/squinky-implicature": 371,
+    "pragmeval/squinky-informativeness": 367,
+    "pragmeval/stac": 392,
+    "pragmeval/switchboard": 346,
+    "pragmeval/verifiability": 369,
+    "privacy-200k-Mistral-Large-3": 5205,
+    "prm800k_dpo/solution": 81,
+    "prm800k_dpo/step": 1184,
+    "probability_words_nli/reasoning_1hop": 5,
+    "probability_words_nli/reasoning_2hop": 13,
+    "procedural-typed-decisions/arithmetic": 4,
+    "procedural-typed-decisions/entity_belief_tracking": 4,
+    "procedural-typed-decisions/event_state_reconstruction": 2,
+    "procedural-typed-decisions/evidence_sufficiency": 38,
+    "procedural-typed-decisions/needle_retrieval": 10,
+    "procedural-typed-decisions/partial_observation_calibration": 455,
+    "procedural-typed-decisions/policy_applicability": 535,
+    "procedural-typed-decisions/policy_under_uncertainty": 28,
+    "procedural-typed-decisions/record_aggregation": 69,
+    "procedural-typed-decisions/state_perturbation": 123,
+    "procedural-typed-decisions/taxonomy_routing": 3,
+    "prompt-injection-dataset": 26,
+    "prompt-injections": 10,
+    "proofwriter": 5322,
+    "propsegment/nli": 1125,
+    "prost": 6,
+    "proto_qa/proto_qa": 9,
+    "qasc": 464,
+    "quail": 10496,
+    "quarel": 1955,
+    "quartz": 144,
+    "race-c": 6696,
+    "race/high": 9541,
+    "race/middle": 9510,
+    "recast/recast_factuality": 104,
+    "recast/recast_megaveridicality": 208,
+    "recast/recast_ner": 47,
+    "recast/recast_puns": 14,
+    "recast/recast_verbcorner": 280,
+    "recast/recast_verbnet": 7,
+    "recast_white/dpr": 1617,
+    "recast_white/fnplus": 1641,
+    "recast_white/sprl": 1655,
+    "reclor": 5180,
+    "regset": 6065,
+    "resnli": 207,
+    "riddle_sense": 4011,
+    "robustLR": 2732,
+    "robust_nli/IS_CS": 12,
+    "robust_nli/LI_LI": 12,
+    "robust_nli/PI_CD": 27,
+    "robust_nli/PI_SP": 8,
+    "robust_nli/ST_LM": 9,
+    "robust_nli/ST_NE": 9,
+    "robust_nli/ST_SE": 19,
+    "robust_nli/ST_WO": 3,
+    "robust_nli_is_sd": 17,
+    "robust_nli_li_ts": 24,
+    "rotten_tomatoes": 3959,
+    "ruletaker": 5,
+    "rumoureval_2019/RumourEval2019": 339,
+    "safe-guard-prompt-injection": 3923,
+    "scicite": 4893,
+    "scientific-exaggeration-detection": 8,
+    "scifact_entailment": 1284,
+    "sciie": 4140,
+    "scinli": 339,
+    "sciq": 6534,
+    "scitail/snli_format": 4210,
+    "scone": 627,
+    "scruples/verdict_votes": 9,
+    "seahorse_summarization_evaluation": 25068,
+    "sem_eval_2010_task_8": 93,
+    "semantic_fragments_nli": 5274,
+    "sen-making/1": 1994,
+    "sen-making/2": 2008,
+    "sharc": 5103,
+    "shell-safety-v2": 62,
+    "sherliic": 91,
+    "sick/label": 4828,
+    "sick/relatedness": 4439,
+    "silicone/dyda_da": 1411,
+    "silicone/maptask": 22,
+    "silicone/oasis": 1461,
+    "silicone/sem": 1280,
+    "simple_pair": 2487,
+    "sms_spam": 45,
+    "snli": 134,
+    "social_i_qa": 346,
+    "spartqa-mchoice": 307,
+    "spartqa-yn": 65,
+    "stackoverflow-questions": 62,
+    "starcon": 24,
+    "stepgame": 279,
+    "strategy-qa": 2640,
+    "sts-companion": 293,
+    "subjectivity": 73,
+    "summarize_from_feedback/comparisons": 47372,
+    "super_glue/axg": 404,
+    "super_glue/boolq": 994,
+    "super_glue/boolq_passage": 1014,
+    "super_glue/cb": 321,
+    "super_glue/copa": 462,
+    "super_glue/multirc": 1054,
+    "super_glue/wic": 1020,
+    "swag/regular": 14295,
+    "syntactic-augmentation-nli": 145,
+    "synthetic-instruct-gptj-pairwise": 8682,
+    "synthetic-retrieval-NLI/binary": 1667,
+    "synthetic-retrieval-NLI/count": 1724,
+    "synthetic-retrieval-NLI/position": 1675,
+    "tasksource_dpo_pairs": 32993,
+    "temporal-nli": 415,
+    "tomi-nli": 241,
+    "toxic-chat/toxicchat0124/jailbreaking": 1979,
+    "toxic-chat/toxicchat0124/toxicity": 1970,
+    "toxic_conversations": 4273,
+    "toxigen-data/annotated": 3908,
+    "tracie": 477,
+    "trec": 6580,
+    "tweet_eval/emoji": 615,
+    "tweet_eval/emotion": 580,
+    "tweet_eval/hate": 610,
+    "tweet_eval/irony": 598,
+    "tweet_eval/offensive": 574,
+    "tweet_eval/sentiment": 528,
+    "tweet_eval/stance_abortion": 584,
+    "tweet_eval/stance_atheism": 530,
+    "tweet_eval/stance_climate": 436,
+    "tweet_eval/stance_feminist": 574,
+    "tweet_eval/stance_hillary": 586,
+    "twentyquestions": 6204,
+    "twitter-financial-news-sentiment": 1,
+    "utilitarianism": 11,
+    "vitaminc": 345,
+    "webgpt_comparisons": 5559,
+    "wice": 279,
+    "wiki_hop/original": 364,
+    "wiki_qa": 4159,
+    "wikimedqa/medwiki": 331,
+    "wildguardmix-cleaned/prompt_harm": 81,
+    "wildguardmix-cleaned/response_harm": 47,
+    "wildguardmix-cleaned/response_refusal": 41,
+    "winodict": 131,
+    "winogrande/winogrande_xl": 12591,
+    "winowhy": 255,
+    "wiqa": 14184,
+    "wnut_17/wnut_17": 82,
+    "wouldyourather": 5,
+    "yahoo_answers_topics": 6263,
+    "yelp_review_full/yelp_review_full": 5161
+  },
+  "files": {
+    "excluded.parquet": {
+      "bytes": 41681395,
+      "sha256": "8edd2abcf71352682d2d5b1a7eda0f67136e3d644060290630155055fe1d8374"
+    },
+    "candidates.parquet": {
+      "bytes": 53743730,
+      "sha256": "6b29c296b6e7e7d78d6ccdf9a50a65e76e864ea511e8affd73886742d1d625c3"
+    }
+  }
+}