MMLU

data

downloadable(HF parquet, cais/mmlu)

license

MIT(OSI/open)

datasheet

present(card documents 57 subjects, splits dev/val/test, schema)

redistributable

yes

paper

public(ICLR 2021)

Clean open data: MIT-licensed, downloadable, with a documented datasheet and a public paper. Fully open.

• https://huggingface.co/datasets/cais/mmlu recorded 2026-09-29

  The dataset page tags the repo "license:mit", its repo state reads "gated":false, and its card says "This is a massive multitask test consisting of multiple-choice questions from various branches of knowledge" with an "auxiliary_train" split listed. Searched the body for non-commercial, noncommercial and holdout: none.

Adoption is measured by Hugging Face downloads of cais/mmlu. MMLU is the canonical general-knowledge benchmark, reported in nearly every frontier model release and still a universal reference despite being largely saturated.

• https://huggingface.co/api/datasets/cais/mmlu recorded 2026-08-12

  477,890 downloads in the trailing 30 days for cais/mmlu

A dataset is not capability-scored, so capability is left unscored. Top models now score between 88% and 94% on MMLU, and the harder MMLU-Pro has become the more discriminating benchmark for comparing frontier models.

• https://huggingface.co/datasets/cais/mmlu recorded 2026-09-29

  The dataset card says "This is a massive multitask test consisting of multiple-choice questions from various branches of knowledge" and the tag list shows "task_ids:multiple-choice-qa". Searched the body for throughput and feature claims for a capability score: none.