Datasets:
License and provenance notice
This compilation uses license: other. No single blanket license is asserted
over all upstream questions, images, or benchmark annotations. The release
schema, cleaning script, and newly authored metadata follow the repository's
project license, but upstream content remains governed by its original terms.
General text
Sources include QANTA, MedQA, GPQA, FinQA, PubMedQA, SuperGLUE, FinanceBench,
PIQA, HellaSwag, SocialIQA, Minerva/Hendrycks mathematics, MMLU-Pro, MathQA,
TACO, APPS, and the recorded long-context sources. The final dataset card must
link each redistributed source to its exact version and license. The historical
wustar/* mirrors recorded in the build config are not anonymously resolvable;
they are provenance identifiers, not evidence of a redistribution license.
Multimodal
MMR-Bench V2 aggregates multiple upstream benchmarks and does not assert a
single license over their questions or images. Preserve the source project's
LICENSES.md and original benchmark terms. The formal test sources include
AI2D, ScienceQA, ScienceQA-IMG, DocVQA, OCR-VQA, and VQAv2/COCO-derived data.
MMR-4000 additionally includes MMStar, RealWorldQA, SEEDBench2 Plus, SimpleVQA,
OCRBench, InfoVQA, ChartQAPro, CharXiv, MathVista, MathVision, MathVerse,
LogicVista, WeMath, MMMU-Pro, BLINK, ERQA, VStarBench, and HallusionBench.
The inherited MMR notice marks binary redistribution as unverified.
Chinese
Sources include AlignBench, C-Eval, CIF-Bench, CLUE-C3, CLUE-CHID, CLUE-CMRC2018, CLUE-CSL, CLUE-OCNLI, CMB-Clin, ChineseSimpleQA, FinEval, FollowBench-zh, GAOKAO-Bench, LawBench, and Chinese LongBench tasks. Preserve the original per-dataset licenses and attribution notices.
Redistribution must be reviewed against each upstream license. When an upstream license does not permit redistribution, publish only IDs, derived routing labels, embeddings where permitted, and a source resolver.