Datasets:
Source attribution and transformations
This collection retains source-specific licenses. It is not uniformly Apache, MIT, or CC0. Each row records the source repository, pinned revision, source row identifier, recorded license, and conversion metadata. Source rights and notices continue to apply to transformed content. The API annotations do not replace those rights.
| Source | Recorded source license | Pinned revision | Training rows |
|---|---|---|---|
| tasksource/ruletaker | Apache-2.0 | a3e0880baeb6ec3d478f4c4d85afe04b21b6cf7f |
209,688 |
| CohereForAI/aya_dataset | Apache-2.0 | f9ea04583f02a8f86404ff6c58bf75fe637df8a2 |
72,934 |
| nyu-mll/multi_nli | Source-specific: CC-BY-3.0; CC-BY-SA-3.0; MIT; see source card | da70db2af9d09693783c3320c4249840212ee221 |
145,870 |
| stanfordnlp/snli | CC-BY-SA-4.0 | cdb5c3d5eed6ead6e5a341c8e56e669bb666725b |
72,934 |
| tals/vitaminc | CC-BY-SA-3.0 | be6febb761b0b2807687e61e0b5282e459df2fa0 |
113,960 |
| google-research-datasets/paws | CC-BY-SA-3.0 (PAWS-Wiki) | 161ece9501cf0a11f3e48bd356eaa82de46d6a09 |
44,672 |
| rajpurkar/squad_v2 | CC-BY-SA-4.0 | 3ffb306f725f7d2ce8394bc1873b24868140c412 |
159,546 |
| nvidia/OpenMathInstruct-2 | CC-BY-4.0 | 469216e3f46f4dacf476b382e192485ea51a143e |
282,626 |
| Lichess/chess-puzzles | CC0-1.0 | 479ea9bc9f681385f5adb23fa27a96c2dc8ae599 |
89,345 |
| BAAI/TACO | Apache-2.0 | d593ed0a2becbbc952230bb89be09189bf1056dc |
84,786 |
| tasksource/bigbench | Apache-2.0; individual tasks retain upstream notices | 210c156767d2f4f05d2f4fd0bb275017a67040fd |
133,988 |
| allenai/tulu-3-sft-mixture | ODC-BY-1.0; underlying source licenses apply | b14afda60f1bbebe55d5d2fa1e4df5042f97f8be |
227,922 |
| google/civil_comments | CC0-1.0 | f2970eb3a55777454c94069077cc8d9b5866312d |
132,195 |
| nvidia/Aegis-AI-Content-Safety-Dataset-2.0 | CC-BY-4.0 | d86bb8bedff51d25ac834ab7838f1cc61acb7a2c |
20,968 |
| prometheus-eval/Feedback-Collection | CC-BY-4.0 | 22a339f961c40b07261bea44ec10849c7440b75f |
88,433 |
| nvidia/HelpSteer2 | CC-BY-4.0 | 990b2711a36180dd19d9c94b8627844866f8982a |
44,672 |
| HuggingFaceH4/ultrafeedback_binarized | MIT | 3949bf5f8c17c394422ccfab0c31ea9c20bdeb85 |
44,672 |
| osunlp/Mind2Web | CC-BY-4.0 | 17ece8eb89862368edc0cc806acee6fca5163474 |
6,175 |
| glaiveai/glaive-function-calling-v2 | Apache-2.0 | e7f4b6456019f5d8bcb991ef0dd67d8ff23221ac |
24,614 |
Transformations
- NLI and rule datasets become explicit entailment, contradiction, unknown, or derivability decisions; PAWS becomes paraphrase classification.
- SQuAD2 becomes answerability and candidate answer selection. Passage groups stay within one split.
- OpenMathInstruct-2 contributes existing augmented math questions and source answers, converted to answer selection or answer verification. Incorrect candidates come from existing compatible source answers. Answers were not independently proved.
- Lichess puzzles become legal UCI move selection after applying the setup move. Source game identities define groups. These are tactical puzzle decisions, not a full chess-strength benchmark.
- TACO contributes existing program input/output cases. This pipeline does not execute submitted code. Obvious constructive-output tasks are filtered, but residual ambiguity is possible.
- BIG-bench training tasks supply native decisions. BBH task families and unsupported image encodings are excluded.
- Tulu 3 and Aya user requests become nine-way task-routing questions. These rows have no independent source gold label. Aya adds multilingual requests.
- Civil Comments becomes a thresholded toxicity decision; Aegis retains source safety labels. These sources contain offensive or harmful text as classification inputs.
- Prometheus, HelpSteer2, and UltraFeedback contribute rubric scoring or response preference decisions. Subjective ratings and source label noise remain.
- Mind2Web contributes text projections of goal, action history, and element candidates. Splits are grouped by website. This is not visual browser perception.
- Glaive function conversations become tool selection or no-tool decisions; only the first function call is used.
Candidate order is deterministically shuffled, with source gold indices and ordinal values remapped. The original source gold remains separate from teacher predictions. See the included conversion code for exact transformations.