Hugging Face's logo

Datasets:
tasksource
/
tasksource-jev-typed-decisions

Tasks:
Zero-Shot Classification
Text Classification
Question Answering
Modalities:
Text
Formats:
parquet
Languages:
English
multilingual
Size:
1M - 10M
Tags:
tasksource
jev
system-one
runtime-defined-decisions
decision-models
multiple-choice
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
2

Add filtered-full train configuration (1,305,478 retained decisions)

#2
by bowang0911 - opened
base: refs/heads/main
←
from: refs/pr/2
Discussion Files changed
+774
-0
Files changed (16) hide show
  1. README.md +40 -0
  2. filtered-full/README.md +105 -0
  3. filtered-full/filtering-report.json +590 -0
  4. filtered-full/train-00000-of-00013.parquet +3 -0
  5. filtered-full/train-00001-of-00013.parquet +3 -0
  6. filtered-full/train-00002-of-00013.parquet +3 -0
  7. filtered-full/train-00003-of-00013.parquet +3 -0
  8. filtered-full/train-00004-of-00013.parquet +3 -0
  9. filtered-full/train-00005-of-00013.parquet +3 -0
  10. filtered-full/train-00006-of-00013.parquet +3 -0
  11. filtered-full/train-00007-of-00013.parquet +3 -0
  12. filtered-full/train-00008-of-00013.parquet +3 -0
  13. filtered-full/train-00009-of-00013.parquet +3 -0
  14. filtered-full/train-00010-of-00013.parquet +3 -0
  15. filtered-full/train-00011-of-00013.parquet +3 -0
  16. filtered-full/train-00012-of-00013.parquet +3 -0
README.md CHANGED
@@ -35,6 +35,10 @@ configs:
35
  path: full/validation-*
36
  - split: test
37
  path: full/test-*
 
 
 
 
38
  dataset_info:
39
  - config_name: default
40
  features:
@@ -120,6 +124,40 @@ dataset_info:
120
  num_examples: 15122
121
  download_size: 1596761311
122
  dataset_size: 3234546031
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
123
  ---
124
 
125
  # tasksource-jev-typed-decisions
@@ -183,6 +221,8 @@ print(row["state"], row["question"], row["options"], row["target"])
183
  - `default`: a steered mix of about 1M train rows. Sources are first gated on label correctness, then weighted by how interesting they are and how close they sit to the zone of proximal development (judged by decision models). Two-option tasks get fewer rows, and procedural generators get 12%. No row is repeated. Validation and test are the full eval splits, restricted to the mixed sources. Buckets and shares are in [`jev_mixes.py`](https://github.com/sileod/tasksource/blob/main/src/tasksource/metadata/jev_mixes.py) and per-source scores in `jev_source_scores.csv`.
184
  - `full`: every row that passed the build, with per-source caps (about 2.5M train rows).
185
 
 
 
186
  ## Format
187
 
188
  | field | meaning |
 
35
  path: full/validation-*
36
  - split: test
37
  path: full/test-*
38
+ - config_name: filtered-full
39
+ data_files:
40
+ - split: train
41
+ path: filtered-full/train-*.parquet
42
  dataset_info:
43
  - config_name: default
44
  features:
 
124
  num_examples: 15122
125
  download_size: 1596761311
126
  dataset_size: 3234546031
127
+ - config_name: filtered-full
128
+ features:
129
+ - name: state
130
+ dtype: string
131
+ - name: kind
132
+ dtype: string
133
+ - name: id
134
+ dtype: string
135
+ - name: options
136
+ list: string
137
+ - name: target
138
+ list: float64
139
+ - name: question
140
+ dtype: string
141
+ - name: source
142
+ dtype: string
143
+ - name: variant
144
+ dtype: string
145
+ - name: split
146
+ dtype: string
147
+ - name: group_id
148
+ dtype: string
149
+ - name: question_id
150
+ dtype: string
151
+ - name: license
152
+ dtype: string
153
+ - name: license_use
154
+ dtype: string
155
+ splits:
156
+ - name: train
157
+ num_bytes: 1633306939
158
+ num_examples: 1305478
159
+ download_size: 335270018
160
+ dataset_size: 1633306939
161
  ---
162
 
163
  # tasksource-jev-typed-decisions
 
221
  - `default`: a steered mix of about 1M train rows. Sources are first gated on label correctness, then weighted by how interesting they are and how close they sit to the zone of proximal development (judged by decision models). Two-option tasks get fewer rows, and procedural generators get 12%. No row is repeated. Validation and test are the full eval splits, restricted to the mixed sources. Buckets and shares are in [`jev_mixes.py`](https://github.com/sileod/tasksource/blob/main/src/tasksource/metadata/jev_mixes.py) and per-source scores in `jev_source_scores.csv`.
222
  - `full`: every row that passed the build, with per-source caps (about 2.5M train rows).
223
 
224
+ - `filtered-full`: 1,305,478 retained `train` decisions from the frozen `full` revision `d2ab1d12be4463fb7ac1f877ed2193b3bb59642a`. Applies commercial-license metadata gating, benchmark-overlap/structural filters, and conservative GLM-5.3-Flash review. Uncertain cases are retained; original row values and soft targets are preserved. No validation or test split is added. This is a filtered historical snapshot and excludes additions made after that revision. See [filtering methods and counts](filtered-full/README.md).
225
+
226
  ## Format
227
 
228
  | field | meaning |
filtered-full/README.md ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # filtered-full
2
+
3
+ `filtered-full` contains **1,305,478 training decisions** retained from the `full/train`
4
+ split of `tasksource/tasksource-jev-typed-decisions` at revision
5
+ [`d2ab1d12be4463fb7ac1f877ed2193b3bb59642a`](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions/tree/d2ab1d12be4463fb7ac1f877ed2193b3bb59642a).
6
+ The cleanup ran October 3–4, 2026. This configuration is a filtered historical
7
+ snapshot; it does not include rows added to the repository after that revision.
8
+
9
+ ```python
10
+ from datasets import load_dataset
11
+
12
+ train = load_dataset(
13
+ "tasksource/tasksource-jev-typed-decisions", "filtered-full", split="train"
14
+ )
15
+ ```
16
+
17
+ Only `train` is supplied. Existing `default` and `full` data, configurations, and
18
+ validation/test splits are unchanged.
19
+
20
+ ## Selection counts
21
+
22
+ | Stage | Decisions |
23
+ |---|---:|
24
+ | Frozen source `full/train` | 2,500,000 |
25
+ | Excluded before LLM review | 1,171,377 |
26
+ | Eligible for LLM review | 1,328,623 |
27
+ | Retained as keep | 1,208,448 |
28
+ | Retained as uncertain | 97,030 |
29
+ | Rejected after review and retention checks | 23,138 |
30
+ | Unresolved API errors, excluded | 7 |
31
+ | **Published training decisions** | **1,305,478** |
32
+
33
+ The run has finished, but seven decisions never obtained a successful review.
34
+ They are excluded; this is not a claim of complete LLM review coverage.
35
+
36
+ ## Preprocessing
37
+
38
+ The selection reflects a specific training-data policy as well as label quality.
39
+ It is not a pure GLM label-error filter over all 2.5 million rows.
40
+
41
+ | Pre-review exclusion reason | Decisions |
42
+ |---|---:|
43
+ | Source `license_use` was not `commercial` | 1,025,115 |
44
+ | Explicit benchmark-source holdout | 122,480 |
45
+ | Positional references in options | 8,108 |
46
+ | Protected exact text match | 8,156 |
47
+ | Protected near text match | 1,526 |
48
+ | Protected contained text match | 3,692 |
49
+ | Group exceeded the prompt token budget | 1,128 |
50
+ | Conflicting targets for an identical decision | 1,172 |
51
+
52
+ These checks operated on source groups. Token filtering used the training prompt
53
+ and tokenizer with an 8,192-token budget, including a 64-token reserve. Text
54
+ overlap checks protected the frozen evaluation panels and Tasksource heldouts
55
+ used in the training workflow. The matching thresholds and explicit source
56
+ holdouts are recorded in [filtering-report.json](filtering-report.json).
57
+ Lexical checks are approximate and do not establish universal benchmark
58
+ decontamination. Preserved license fields describe the upstream metadata; this
59
+ configuration does not introduce a new license or replace source terms.
60
+
61
+ ## GLM review
62
+
63
+ The reviewer was **GLM-5.3-Flash**, model identifier `glm_5_3_flash`, accessed
64
+ through llm-api with 768 workers. Review considered the input, question, ordered
65
+ options, and original target distribution. It sought demonstrable defects rather
66
+ than replacing source labels with a model's preferred answer.
67
+
68
+ An initial reviewer and a reasoning-enabled verifier both had to demonstrate an
69
+ obvious defect and recommend rejection. Uncertain cases were retained. Final
70
+ conservative retention checks also restored 1,221 neutral/non-entailment
71
+ disagreements, 89 emotion-classification disagreements, and one ambiguous
72
+ title-versus-procedural-advice case. These 1,311 cases are included in the final
73
+ retained counts above.
74
+
75
+ Identical rows already reviewed during the earlier `default`-configuration
76
+ cleanup reused their review evidence after matching the complete serialized
77
+ review payload. Remaining rows were reviewed separately. Inherited reviews keep
78
+ their original prompt provenance; the later full review added a clarification
79
+ that the state is the input. This is an LLM-assisted cleanup, not human-certified
80
+ ground truth. Model disagreement alone is not proof of an incorrect label.
81
+
82
+ ## Row preservation and export validation
83
+
84
+ The exported Parquet files select original rows by upstream `id` from the frozen
85
+ source files. They preserve all **13 columns present in that historical revision**,
86
+ including text, option order, `kind`, IDs, soft targets, and license metadata.
87
+ Newer repository revisions added an `example_id` column; it is not synthesized
88
+ for this historical configuration.
89
+
90
+ The intermediate training representation used numeric option keys and mapped
91
+ `score` requests to its internal choice interface. The published dataset restores
92
+ the original upstream schema and values. For 7082 retained rows
93
+ whose option wording had been normalized for training, export also restores the
94
+ original option strings. No labels or probabilities are rewritten.
95
+
96
+ Preprocessing retained or removed groups, while the final LLM selection acts on
97
+ individual decisions. Consequently, some `group_id`s contain fewer decisions
98
+ than in the original snapshot. Original row order is retained; do not assume
99
+ every surviving group is contiguous. Group by both `group_id` and `state` when
100
+ reconstructing related requests.
101
+
102
+ Validation checks the frozen input hashes, unique IDs, every retained row against
103
+ the cleanup selection, and every exported value after reading all Parquet shards
104
+ back. Counts, source distributions, selection fingerprints, and shard checksums
105
+ are available in [filtering-report.json](filtering-report.json).
filtered-full/filtering-report.json ADDED
@@ -0,0 +1,590 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "tasksource/tasksource-jev-typed-decisions",
3
+ "configuration": "filtered-full",
4
+ "split": "train",
5
+ "source_config": "full",
6
+ "source_revision": "d2ab1d12be4463fb7ac1f877ed2193b3bb59642a",
7
+ "base_repository_revision": "47f1e1855838f92be44bd1f1689a826a8c3c3a0b",
8
+ "source_rows": 2500000,
9
+ "eligible_rows": 1328623,
10
+ "retained_rows": 1305478,
11
+ "prefilter_exclusions": {
12
+ "license_not_commercial": 1025115,
13
+ "benchmark_source_holdout": 122480,
14
+ "quality_positional_option_reference": 8108,
15
+ "protected_exact_text": 8156,
16
+ "protected_near_text": 1526,
17
+ "protected_contained_text": 3692,
18
+ "group_over_token_limit": 1128,
19
+ "conflicting_exact_decision_targets": 1172
20
+ },
21
+ "llm_rejected_rows": 23138,
22
+ "unresolved_api_error_rows_excluded": 7,
23
+ "retained_keep_rows": 1208448,
24
+ "retained_uncertain_rows": 97030,
25
+ "conservative_retention_overrides": {
26
+ "neutral_or_non_entailment_disagreement_preserved": 1221,
27
+ "emotion_classification_disagreement_preserved": 89,
28
+ "ambiguous_title_vs_procedural_advice_preserved": 1
29
+ },
30
+ "review_model": "glm_5_3_flash",
31
+ "workers": 768,
32
+ "license_use_policy": [
33
+ "commercial"
34
+ ],
35
+ "near_match_policy": "NFKC casefold; five-word shingles for >=20-word text parts; bottom-eight protected anchors probed by all query shingles; verify Jaccard >=0.85 or protected containment >=0.95 with >=20 distinct shingles.",
36
+ "limitations": [
37
+ "Lexical candidate retrieval is approximate; paraphrases/translations can remain.",
38
+ "Shared contexts are conservatively quarantined even when questions differ.",
39
+ "Short states require matching instructions and option texts.",
40
+ "Only the frozen local evaluation panels and Tasksource heldouts were indexed, not every upstream evaluation example.",
41
+ "Source holdouts are explicit aliases, not a proof that all derived-source aliases were enumerated.",
42
+ "Prior CE536 exposure and base-model pretraining contamination cannot be reversed by this filtering."
43
+ ],
44
+ "source_file_sha256": {
45
+ "full/train-00000-of-00013.parquet": "4fbd088cce695b14a8bc223a8b246822bc01771f2c5c5103efbcfd4274aeddcc",
46
+ "full/train-00001-of-00013.parquet": "815353f5aff56164d51957688713d8411b939c1fa18eeccef68d04b5c000c715",
47
+ "full/train-00002-of-00013.parquet": "14daac006804bf1700bc622ccb055da9fc811a419a3d0a5895e33d3c5fe6c992",
48
+ "full/train-00003-of-00013.parquet": "869f351889aad611bcefb7f81a03cc8192cef33de2f4fee9ef045c596f47c1ba",
49
+ "full/train-00004-of-00013.parquet": "631bb0dcb6be88f67c86db7b873d606b596f8c8e8a5378757e76e93cd51d41e7",
50
+ "full/train-00005-of-00013.parquet": "3f93c866f0cc97e1e8deb5d15b0db55c12a1fb6c58fad3194ff18dcea0b03e3c",
51
+ "full/train-00006-of-00013.parquet": "155bcc3dbf05f696bced9294f545682c15d7e2b894cb074c9ebfcb2ec7193720",
52
+ "full/train-00007-of-00013.parquet": "8be4ec6304e1ceb5d414df17fda3f8ca026a39fe0861eebd44b258dc293ab26e",
53
+ "full/train-00008-of-00013.parquet": "23ee32e6b3789506580b1a8bb9aa2bbe5baff5fe39ff0cb2f68290fac7075dba",
54
+ "full/train-00009-of-00013.parquet": "d89249adc59f1bf988dce946ac6e68c56b2eb8f0957f0f0afba967b5e1a2b37e",
55
+ "full/train-00010-of-00013.parquet": "8cc14c1dc5db8facdb3b994ebf9c05f235693b2e085a637a76ba655a4c0df0f4",
56
+ "full/train-00011-of-00013.parquet": "6d85929919f3c353c02ee9178f89a2ca56723816fac9ef93d1643f9d03917d5e",
57
+ "full/train-00012-of-00013.parquet": "0a7ad9e91130d379205a8d74dc79f42654324876e59abe140928c565fc18ab89"
58
+ },
59
+ "selection_artifact_sha256": "917d09d70d34131271f3721ccf50466539bc86077a9b21ac45db04ddc84f0ffb",
60
+ "retained_id_sha256": "6a85e41b1df79781cfc9f1e520c5629819228054d3deaaa8835e06716f68e109",
61
+ "retained_content_sha256": "1be7e0176ce404d9d3985bf12a25dfc9e52b6d121f1143cc5e70241335e7f0d4",
62
+ "soft_target_rows": 137840,
63
+ "normalized_option_rows_restored_to_original": 7082,
64
+ "counts": {
65
+ "source": {
66
+ "CONDAQA": 4378,
67
+ "CREAK": 3890,
68
+ "FLUTE": 3893,
69
+ "FOL-nli": 13594,
70
+ "HatemojiBuild": 3480,
71
+ "HelpSteer/coherence": 948,
72
+ "HelpSteer/complexity": 938,
73
+ "HelpSteer/correctness": 930,
74
+ "HelpSteer/helpfulness": 930,
75
+ "HelpSteer/verbosity": 910,
76
+ "HelpSteer2/coherence": 1244,
77
+ "HelpSteer2/complexity": 1270,
78
+ "HelpSteer2/correctness": 1246,
79
+ "HelpSteer2/helpfulness": 1256,
80
+ "HelpSteer2/verbosity": 1252,
81
+ "HelpSteer3/edit_quality": 3413,
82
+ "HelpSteer3/feedback": 1933,
83
+ "HelpSteer3/preference": 9247,
84
+ "HelpSteer3/preference_strength": 1909,
85
+ "HelpSteer3/principle": 1077,
86
+ "I2D2": 3855,
87
+ "MSciNLI": 5013,
88
+ "MedQA-USMLE-4-options-hf": 10295,
89
+ "PARADISE": 14156,
90
+ "PARARULE-Plus": 4334,
91
+ "Prompt-injection-dataset/full": 3938,
92
+ "PromptShield": 3912,
93
+ "SDOH-NLI": 2269,
94
+ "SpaceNLI": 5123,
95
+ "Touche23-ValueEval": 2005,
96
+ "TroFi": 2035,
97
+ "TuringBench": 9064,
98
+ "UNLI": 29967,
99
+ "UltraFeedback-paired": 11986,
100
+ "WANLI": 14136,
101
+ "acceptability-prediction/binary_votes": 2220,
102
+ "acceptability-prediction/rating_votes": 4245,
103
+ "agent_action_safety": 2264,
104
+ "amazon_counterfactual/en": 3713,
105
+ "amazon_polarity/amazon_polarity": 4281,
106
+ "arct": 1326,
107
+ "args_me": 4195,
108
+ "autotnli": 4165,
109
+ "babi_nli/basic-coreference": 332,
110
+ "babi_nli/basic-deduction": 320,
111
+ "babi_nli/basic-induction": 315,
112
+ "babi_nli/compound-coreference": 315,
113
+ "babi_nli/conjunction": 323,
114
+ "babi_nli/counting": 304,
115
+ "babi_nli/indefinite-knowledge": 325,
116
+ "babi_nli/lists-sets": 325,
117
+ "babi_nli/path-finding": 307,
118
+ "babi_nli/positional-reasoning": 295,
119
+ "babi_nli/simple-negation": 324,
120
+ "babi_nli/single-supporting-fact": 294,
121
+ "babi_nli/size-reasoning": 317,
122
+ "babi_nli/three-arg-relations": 330,
123
+ "babi_nli/three-supporting-facts": 324,
124
+ "babi_nli/time-reasoning": 319,
125
+ "babi_nli/two-arg-relations": 317,
126
+ "babi_nli/two-supporting-facts": 298,
127
+ "babi_nli/yes-no-questions": 328,
128
+ "balanced-copa": 1006,
129
+ "blog_authorship_corpus/age": 2055,
130
+ "blog_authorship_corpus/gender": 2050,
131
+ "blog_authorship_corpus/job": 2088,
132
+ "breaking_nli": 4721,
133
+ "cicero": 13237,
134
+ "civil_comments/identity_attack_share": 8643,
135
+ "civil_comments/insult_share": 8544,
136
+ "civil_comments/obscene_share": 8646,
137
+ "civil_comments/severe_toxicity_share": 8653,
138
+ "civil_comments/sexual_explicit_share": 8649,
139
+ "civil_comments/threat_share": 8641,
140
+ "civil_comments/toxicity_share": 8643,
141
+ "clcd-english": 3782,
142
+ "cloth": 14064,
143
+ "codah/codah": 2839,
144
+ "commonsense_qa_2.0": 3893,
145
+ "conceptrules_v2": 6143,
146
+ "corr2cause": 4996,
147
+ "cosmos_qa": 13092,
148
+ "cycic_classification": 3839,
149
+ "cycic_multiplechoice": 7254,
150
+ "dbpedia_14/dbpedia_14": 3359,
151
+ "defeasible-nli/atomic": 1835,
152
+ "defeasible-nli/snli": 1810,
153
+ "defeasible-nli/social": 1805,
154
+ "dgen": 2359,
155
+ "discovery/discovery": 6967,
156
+ "disrpt/eng.dep.scidtb.rels": 6705,
157
+ "dnd_style_intents": 6214,
158
+ "doc-nli": 23729,
159
+ "docred": 10233,
160
+ "dynahate": 4346,
161
+ "dynasent/r1_votes": 25149,
162
+ "dynasent/r2_votes": 13625,
163
+ "e-CARE": 9339,
164
+ "emo/emo2019": 2868,
165
+ "equate": 2145,
166
+ "ethics/commonsense": 1313,
167
+ "ethics/deontology": 1310,
168
+ "ethics/justice": 1304,
169
+ "ethics/virtue": 1320,
170
+ "ethos/multilabel": 616,
171
+ "feasibilityQA": 3688,
172
+ "fever-evidence-related": 2535,
173
+ "fig-qa": 8539,
174
+ "folio": 1126,
175
+ "fool-me-twice": 3888,
176
+ "gen_debiased_nli/mnli_par_z": 626,
177
+ "gen_debiased_nli/mnli_seq_z": 614,
178
+ "gen_debiased_nli/mnli_z_aug": 605,
179
+ "gen_debiased_nli/snli_par_z": 614,
180
+ "gen_debiased_nli/snli_seq_z": 588,
181
+ "gen_debiased_nli/snli_z_aug": 594,
182
+ "github-issue-similarity": 3976,
183
+ "go_emotions/simplified": 3887,
184
+ "goal-step-wikihow/goal": 8376,
185
+ "goal-step-wikihow/order": 8466,
186
+ "goal-step-wikihow/step": 8277,
187
+ "google_wellformed_query": 8161,
188
+ "hate_speech18": 3975,
189
+ "head_qa/en": 2999,
190
+ "headline_cause/en_simple": 4771,
191
+ "health_fact": 5144,
192
+ "help-nli": 5484,
193
+ "hh-rlhf/harmless-base": 6499,
194
+ "hh-rlhf/helpful-base": 6473,
195
+ "hh-rlhf/helpful-online": 6482,
196
+ "hh-rlhf/helpful-rejection-sampled": 6524,
197
+ "hlgd": 3045,
198
+ "hope_edi/english": 5055,
199
+ "hyperpartisan_news": 629,
200
+ "implicatures": 1024,
201
+ "jigsaw_toxicity": 4271,
202
+ "lex_glue/case_hold": 13906,
203
+ "lex_glue/ledgar": 2331,
204
+ "lex_glue/scotus": 1417,
205
+ "lex_glue/unfair_tos": 2295,
206
+ "logical-entailment": 4255,
207
+ "lonli": 5077,
208
+ "math_qa": 11992,
209
+ "medmcqa": 13674,
210
+ "mindgames": 3282,
211
+ "monotonicity-entailment": 3704,
212
+ "moral_stories/full": 8862,
213
+ "multilingual/AfriSenti-twitter-sentiment/amh": 464,
214
+ "multilingual/AfriSenti-twitter-sentiment/arq": 470,
215
+ "multilingual/AfriSenti-twitter-sentiment/ary": 476,
216
+ "multilingual/AfriSenti-twitter-sentiment/hau": 485,
217
+ "multilingual/AfriSenti-twitter-sentiment/ibo": 463,
218
+ "multilingual/AfriSenti-twitter-sentiment/kin": 464,
219
+ "multilingual/AfriSenti-twitter-sentiment/pcm": 465,
220
+ "multilingual/AfriSenti-twitter-sentiment/por": 455,
221
+ "multilingual/AfriSenti-twitter-sentiment/swa": 462,
222
+ "multilingual/AfriSenti-twitter-sentiment/tso": 470,
223
+ "multilingual/AfriSenti-twitter-sentiment/twi": 464,
224
+ "multilingual/AfriSenti-twitter-sentiment/yor": 463,
225
+ "multilingual/MLMA_hate_speech": 3080,
226
+ "multilingual/NusaX-senti/ace": 398,
227
+ "multilingual/NusaX-senti/ban": 394,
228
+ "multilingual/NusaX-senti/bbc": 394,
229
+ "multilingual/NusaX-senti/bjn": 403,
230
+ "multilingual/NusaX-senti/bug": 406,
231
+ "multilingual/NusaX-senti/eng": 393,
232
+ "multilingual/NusaX-senti/ind": 404,
233
+ "multilingual/NusaX-senti/jav": 410,
234
+ "multilingual/NusaX-senti/mad": 401,
235
+ "multilingual/NusaX-senti/min": 412,
236
+ "multilingual/NusaX-senti/nij": 409,
237
+ "multilingual/NusaX-senti/sun": 400,
238
+ "multilingual/americas_nli/all_languages": 4107,
239
+ "multilingual/disrpt/deu.rst.pcc.rels": 762,
240
+ "multilingual/disrpt/eus.rst.ert.rels": 768,
241
+ "multilingual/disrpt/fas.rst.prstc.rels": 760,
242
+ "multilingual/disrpt/fra.sdrt.annodis.rels": 777,
243
+ "multilingual/disrpt/nld.rst.nldt.rels": 762,
244
+ "multilingual/disrpt/por.rst.cstn.rels": 773,
245
+ "multilingual/disrpt/rus.rst.rrt.rels": 744,
246
+ "multilingual/disrpt/spa.rst.rststb.rels": 775,
247
+ "multilingual/disrpt/tha.pdtb.tdtb.rels": 779,
248
+ "multilingual/disrpt/zho.rst.gcdt.rels": 732,
249
+ "multilingual/exams/multilingual": 8639,
250
+ "multilingual/klue/nli": 1946,
251
+ "multilingual/klue/sts": 1993,
252
+ "multilingual/klue/ynat": 1909,
253
+ "multilingual/masakhanews/amh": 325,
254
+ "multilingual/masakhanews/eng": 328,
255
+ "multilingual/masakhanews/fra": 331,
256
+ "multilingual/masakhanews/hau": 333,
257
+ "multilingual/masakhanews/ibo": 329,
258
+ "multilingual/masakhanews/lin": 325,
259
+ "multilingual/masakhanews/lug": 320,
260
+ "multilingual/masakhanews/orm": 325,
261
+ "multilingual/masakhanews/pcm": 322,
262
+ "multilingual/masakhanews/run": 306,
263
+ "multilingual/masakhanews/sna": 312,
264
+ "multilingual/masakhanews/som": 320,
265
+ "multilingual/masakhanews/swa": 314,
266
+ "multilingual/masakhanews/tir": 327,
267
+ "multilingual/masakhanews/xho": 315,
268
+ "multilingual/masakhanews/yor": 308,
269
+ "multilingual/massive": 6631,
270
+ "multilingual/miam": 5703,
271
+ "multilingual/mtop": 6988,
272
+ "multilingual/multilingual-sentiments/all": 4961,
273
+ "multilingual/offenseval_dravidian/kannada": 2174,
274
+ "multilingual/offenseval_dravidian/malayalam": 2169,
275
+ "multilingual/offenseval_dravidian/tamil": 2192,
276
+ "multilingual/tweet_sentiment_multilingual": 4946,
277
+ "multilingual/universal-joy": 5246,
278
+ "multilingual/x-stance": 3181,
279
+ "multilingual/xcopa/et": 54,
280
+ "multilingual/xcopa/ht": 57,
281
+ "multilingual/xcopa/id": 63,
282
+ "multilingual/xcopa/it": 53,
283
+ "multilingual/xcopa/qu": 57,
284
+ "multilingual/xcopa/sw": 57,
285
+ "multilingual/xcopa/ta": 55,
286
+ "multilingual/xcopa/th": 54,
287
+ "multilingual/xcopa/tr": 55,
288
+ "multilingual/xcopa/translation-et": 54,
289
+ "multilingual/xcopa/translation-ht": 56,
290
+ "multilingual/xcopa/translation-id": 56,
291
+ "multilingual/xcopa/translation-it": 55,
292
+ "multilingual/xcopa/translation-sw": 56,
293
+ "multilingual/xcopa/translation-ta": 58,
294
+ "multilingual/xcopa/translation-th": 54,
295
+ "multilingual/xcopa/translation-tr": 53,
296
+ "multilingual/xcopa/translation-vi": 56,
297
+ "multilingual/xcopa/translation-zh": 58,
298
+ "multilingual/xcopa/vi": 58,
299
+ "multilingual/xcopa/zh": 58,
300
+ "multilingual/xcsr/X-CODAH-ar": 159,
301
+ "multilingual/xcsr/X-CODAH-de": 164,
302
+ "multilingual/xcsr/X-CODAH-en": 166,
303
+ "multilingual/xcsr/X-CODAH-es": 161,
304
+ "multilingual/xcsr/X-CODAH-fr": 167,
305
+ "multilingual/xcsr/X-CODAH-hi": 162,
306
+ "multilingual/xcsr/X-CODAH-it": 165,
307
+ "multilingual/xcsr/X-CODAH-jap": 166,
308
+ "multilingual/xcsr/X-CODAH-nl": 163,
309
+ "multilingual/xcsr/X-CODAH-pl": 156,
310
+ "multilingual/xcsr/X-CODAH-pt": 163,
311
+ "multilingual/xcsr/X-CODAH-ru": 164,
312
+ "multilingual/xcsr/X-CODAH-sw": 166,
313
+ "multilingual/xcsr/X-CODAH-ur": 168,
314
+ "multilingual/xcsr/X-CODAH-vi": 163,
315
+ "multilingual/xcsr/X-CODAH-zh": 162,
316
+ "multilingual/xstory_cloze/ar": 362,
317
+ "multilingual/xstory_cloze/en": 367,
318
+ "multilingual/xstory_cloze/es": 351,
319
+ "multilingual/xstory_cloze/eu": 361,
320
+ "multilingual/xstory_cloze/hi": 364,
321
+ "multilingual/xstory_cloze/id": 354,
322
+ "multilingual/xstory_cloze/my": 362,
323
+ "multilingual/xstory_cloze/ru": 352,
324
+ "multilingual/xstory_cloze/sw": 356,
325
+ "multilingual/xstory_cloze/te": 351,
326
+ "multilingual/xstory_cloze/zh": 352,
327
+ "nan-nli": 212,
328
+ "naturallogic": 4363,
329
+ "numer_sense": 5721,
330
+ "onestop_qa": 1444,
331
+ "open_question_type": 4906,
332
+ "patent-phrase-similarity": 5954,
333
+ "poem_sentiment": 1154,
334
+ "prm800k_dpo/solution": 4024,
335
+ "prm800k_dpo/step": 2258,
336
+ "probability_words_nli/reasoning_1hop": 1972,
337
+ "probability_words_nli/reasoning_2hop": 1957,
338
+ "procedural-typed-decisions/arithmetic": 19995,
339
+ "procedural-typed-decisions/entity_belief_tracking": 19996,
340
+ "procedural-typed-decisions/event_state_reconstruction": 19996,
341
+ "procedural-typed-decisions/evidence_sufficiency": 19963,
342
+ "procedural-typed-decisions/multi_view_adjudication": 19998,
343
+ "procedural-typed-decisions/needle_retrieval": 19988,
344
+ "procedural-typed-decisions/partial_observation_calibration": 19545,
345
+ "procedural-typed-decisions/policy_applicability": 19463,
346
+ "procedural-typed-decisions/policy_under_uncertainty": 19973,
347
+ "procedural-typed-decisions/record_aggregation": 19931,
348
+ "procedural-typed-decisions/state_perturbation": 19875,
349
+ "procedural-typed-decisions/table_lookup": 10004,
350
+ "procedural-typed-decisions/taxonomy_routing": 20001,
351
+ "prompt-injection-dataset": 3968,
352
+ "prompt-injections": 597,
353
+ "propsegment/nli": 4079,
354
+ "prost": 7491,
355
+ "proto_qa/proto_qa": 7893,
356
+ "qasc": 6233,
357
+ "quartz": 2991,
358
+ "recast/recast_factuality": 737,
359
+ "recast/recast_megaveridicality": 668,
360
+ "recast/recast_ner": 791,
361
+ "recast/recast_puns": 822,
362
+ "recast/recast_sentiment": 860,
363
+ "recast/recast_verbcorner": 555,
364
+ "recast/recast_verbnet": 844,
365
+ "resnli": 4076,
366
+ "robust_nli/IS_CS": 195,
367
+ "robust_nli/LI_LI": 214,
368
+ "robust_nli/PI_CD": 183,
369
+ "robust_nli/PI_SP": 201,
370
+ "robust_nli/ST_LM": 207,
371
+ "robust_nli/ST_NE": 209,
372
+ "robust_nli/ST_SE": 180,
373
+ "robust_nli/ST_WO": 192,
374
+ "robust_nli_is_sd": 889,
375
+ "robust_nli_li_ts": 564,
376
+ "ruletaker": 4331,
377
+ "rumoureval_2019/RumourEval2019": 4499,
378
+ "scientific-exaggeration-detection": 116,
379
+ "scinli": 4978,
380
+ "scone": 3265,
381
+ "scruples/verdict_votes": 14726,
382
+ "sem_eval_2010_task_8": 5731,
383
+ "shell-safety-v2": 5216,
384
+ "sherliic": 1071,
385
+ "silicone/maptask": 1411,
386
+ "sms_spam": 3844,
387
+ "snips_built_in_intents": 396,
388
+ "snli": 5151,
389
+ "social_i_qa": 13935,
390
+ "spartqa-mchoice": 16888,
391
+ "spartqa-yn": 7506,
392
+ "stackoverflow-questions": 5126,
393
+ "starcon": 1716,
394
+ "stepgame": 5933,
395
+ "sts-companion": 4467,
396
+ "subjectivity": 836,
397
+ "syntactic-augmentation-nli": 4809,
398
+ "temporal-nli": 4863,
399
+ "tomi-nli": 5542,
400
+ "tracie": 878,
401
+ "tweets_hate_speech_detection": 4311,
402
+ "twitter-financial-news-sentiment": 4920,
403
+ "utilitarianism": 2128,
404
+ "vitaminc": 4982,
405
+ "wice": 3881,
406
+ "wiki_hop/original": 15572,
407
+ "wikimedqa/medwiki": 15571,
408
+ "wildguardmix-cleaned/prompt_harm": 1612,
409
+ "wildguardmix-cleaned/response_harm": 1802,
410
+ "wildguardmix-cleaned/response_refusal": 1711,
411
+ "winodict": 1410,
412
+ "winowhy": 3030,
413
+ "wnut_17/wnut_17": 7361,
414
+ "wouldyourather": 1029
415
+ },
416
+ "kind": {
417
+ "choice": 947205,
418
+ "noul": 272281,
419
+ "score": 85992
420
+ },
421
+ "variant": {
422
+ "criteria_permutation": 35879,
423
+ "direct": 1147750,
424
+ "instruction_paraphrase": 20579,
425
+ "label_verification": 35933,
426
+ "packed_derived": 52708,
427
+ "paired_text_format": 12629
428
+ },
429
+ "license_use": {
430
+ "commercial": 1305478
431
+ }
432
+ },
433
+ "shards": [
434
+ {
435
+ "path": "filtered-full/train-00000-of-00013.parquet",
436
+ "rows": 100226,
437
+ "bytes": 25731157,
438
+ "sha256": "886002369ca5d60cc57f02a69a6d023d23a208e50b534ff8cc6b8af952d0ee25"
439
+ },
440
+ {
441
+ "path": "filtered-full/train-00001-of-00013.parquet",
442
+ "rows": 100492,
443
+ "bytes": 25682657,
444
+ "sha256": "b607dffebc51cce8f20c2fd91814d1e586ce97464609eaddad0707250d2debb4"
445
+ },
446
+ {
447
+ "path": "filtered-full/train-00002-of-00013.parquet",
448
+ "rows": 100946,
449
+ "bytes": 26261894,
450
+ "sha256": "653af83130e8975d226b9cf96c527870deea9cb2c3d89000229d71be1eddade1"
451
+ },
452
+ {
453
+ "path": "filtered-full/train-00003-of-00013.parquet",
454
+ "rows": 99945,
455
+ "bytes": 25551178,
456
+ "sha256": "9b18d075997f68b7f21d8a165d31ee975d8db1662ba8a93421157ef4d355b11a"
457
+ },
458
+ {
459
+ "path": "filtered-full/train-00004-of-00013.parquet",
460
+ "rows": 100579,
461
+ "bytes": 25718485,
462
+ "sha256": "293c3a5d56b3f1457e7496833095f7d6f326be29bec3c8c1518e408a4f108888"
463
+ },
464
+ {
465
+ "path": "filtered-full/train-00005-of-00013.parquet",
466
+ "rows": 100081,
467
+ "bytes": 25686366,
468
+ "sha256": "41debd0ab9536f4f8b088416e3e4ceea5885edb33e56bdc815f3d776c5e4056e"
469
+ },
470
+ {
471
+ "path": "filtered-full/train-00006-of-00013.parquet",
472
+ "rows": 100649,
473
+ "bytes": 25490143,
474
+ "sha256": "7a1f543fa82defc0511ce9fa02493172c273585233a44d8622badcbd0d9b5475"
475
+ },
476
+ {
477
+ "path": "filtered-full/train-00007-of-00013.parquet",
478
+ "rows": 100537,
479
+ "bytes": 25660020,
480
+ "sha256": "d89d241dac8edb39db2ca78a6a36348f9f9999c40998b2515f918f00465542c7"
481
+ },
482
+ {
483
+ "path": "filtered-full/train-00008-of-00013.parquet",
484
+ "rows": 100335,
485
+ "bytes": 25930607,
486
+ "sha256": "70a784a11782a0200e588d84e35cb97ab816bbdee566472b2294ca8cc7ec8d03"
487
+ },
488
+ {
489
+ "path": "filtered-full/train-00009-of-00013.parquet",
490
+ "rows": 100460,
491
+ "bytes": 25873105,
492
+ "sha256": "e8a0d74015b82d2d7f30cbb5c8018d717be8da5d00cecb6661369de9269aa7be"
493
+ },
494
+ {
495
+ "path": "filtered-full/train-00010-of-00013.parquet",
496
+ "rows": 100565,
497
+ "bytes": 25764127,
498
+ "sha256": "bd01d8f741169b42bed4095b5dbae85f8d345c1a92b02f60fcbedc0ea8b2a6c1"
499
+ },
500
+ {
501
+ "path": "filtered-full/train-00011-of-00013.parquet",
502
+ "rows": 100405,
503
+ "bytes": 25798190,
504
+ "sha256": "b7601e3d6114481252bc5b5b60567a48b91a9f0490caf4c24a16384d11f5c8b5"
505
+ },
506
+ {
507
+ "path": "filtered-full/train-00012-of-00013.parquet",
508
+ "rows": 100258,
509
+ "bytes": 26122089,
510
+ "sha256": "937b49820b25257e50520c8a25f2f7257c0e643532718f6d07d1926897efd922"
511
+ }
512
+ ],
513
+ "row_order": "Original frozen full/train shard order, filtered by retained upstream ID.",
514
+ "validation": {
515
+ "unique_source_ids": true,
516
+ "every_retained_row_matches_selection": true,
517
+ "all_exported_values_roundtrip_verified": true,
518
+ "no_relabeling": true,
519
+ "original_configs_unchanged": true,
520
+ "only_train_split_added": true
521
+ },
522
+ "reused_review_rows": 541517,
523
+ "rows_requiring_new_review": 787106,
524
+ "benchmark_source_holdout_rule": {
525
+ "blocked_roots": [
526
+ "acos",
527
+ "ai2_arc",
528
+ "anli",
529
+ "api-bank",
530
+ "arc-challenge",
531
+ "arc-easy",
532
+ "banking77",
533
+ "bbh",
534
+ "belebele",
535
+ "bfcl",
536
+ "bigbench",
537
+ "blimp",
538
+ "bright",
539
+ "chessbench",
540
+ "circa",
541
+ "cladder",
542
+ "clinc150",
543
+ "clinc_oos",
544
+ "commonsense_qa",
545
+ "contract-nli",
546
+ "contractnli",
547
+ "cos_e",
548
+ "cruxeval",
549
+ "esci",
550
+ "financial_phrasebank",
551
+ "finentity",
552
+ "forecastbench",
553
+ "gpqa",
554
+ "gsm8k",
555
+ "hellaswag",
556
+ "hle",
557
+ "hover",
558
+ "hover-3way",
559
+ "humicroedit",
560
+ "isarcasmeval",
561
+ "it-support-tickets",
562
+ "jevbench",
563
+ "judgebench",
564
+ "mmlu",
565
+ "mmlu-pro",
566
+ "musr",
567
+ "nli4ct_semeval2024",
568
+ "openbookqa",
569
+ "phishnchips",
570
+ "ragtruth",
571
+ "routerbench",
572
+ "sata-bench",
573
+ "schema_guided_dstc8",
574
+ "sgd",
575
+ "sgd-x",
576
+ "simplebench",
577
+ "tabfact",
578
+ "toolret",
579
+ "truthfulqa",
580
+ "vast",
581
+ "when2call",
582
+ "winogrande"
583
+ ],
584
+ "root_matching": "Take the part before the first slash, casefold, then remove non-alphanumeric ASCII characters before comparing.",
585
+ "casefolded_source_prefixes": [
586
+ "multilingual/xcsr/x-csqa-"
587
+ ],
588
+ "implementation_sha256": "fa3ed2826c1d9b0a40f6cc4102ebd46c2609ba8100b0d1d80be9293764cabe57"
589
+ }
590
+ }
filtered-full/train-00000-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:886002369ca5d60cc57f02a69a6d023d23a208e50b534ff8cc6b8af952d0ee25
3
+ size 25731157
filtered-full/train-00001-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b607dffebc51cce8f20c2fd91814d1e586ce97464609eaddad0707250d2debb4
3
+ size 25682657
filtered-full/train-00002-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:653af83130e8975d226b9cf96c527870deea9cb2c3d89000229d71be1eddade1
3
+ size 26261894
filtered-full/train-00003-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9b18d075997f68b7f21d8a165d31ee975d8db1662ba8a93421157ef4d355b11a
3
+ size 25551178
filtered-full/train-00004-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:293c3a5d56b3f1457e7496833095f7d6f326be29bec3c8c1518e408a4f108888
3
+ size 25718485
filtered-full/train-00005-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:41debd0ab9536f4f8b088416e3e4ceea5885edb33e56bdc815f3d776c5e4056e
3
+ size 25686366
filtered-full/train-00006-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7a1f543fa82defc0511ce9fa02493172c273585233a44d8622badcbd0d9b5475
3
+ size 25490143
filtered-full/train-00007-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d89d241dac8edb39db2ca78a6a36348f9f9999c40998b2515f918f00465542c7
3
+ size 25660020
filtered-full/train-00008-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70a784a11782a0200e588d84e35cb97ab816bbdee566472b2294ca8cc7ec8d03
3
+ size 25930607
filtered-full/train-00009-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8a0d74015b82d2d7f30cbb5c8018d717be8da5d00cecb6661369de9269aa7be
3
+ size 25873105
filtered-full/train-00010-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bd01d8f741169b42bed4095b5dbae85f8d345c1a92b02f60fcbedc0ea8b2a6c1
3
+ size 25764127
filtered-full/train-00011-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7601e3d6114481252bc5b5b60567a48b91a9f0490caf4c24a16384d11f5c8b5
3
+ size 25798190
filtered-full/train-00012-of-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:937b49820b25257e50520c8a25f2f7257c0e643532718f6d07d1926897efd922
3
+ size 26122089