MMLU (Massive Multitask Language Understanding) — LLM Evaluation & Benchmarks Dataset
The gold-standard multi-subject benchmark: 16k multiple-choice questions spanning 57 subjects from STEM and humanities to law, medicine, and social sciences. Every major LLM is evaluated on MMLU. A model scoring >80% is considered strong; GPT-4 scores ~87%, Llama 3 70B ~82%.
Dataset Details
| Provider | cais |
| Category | Evaluation & Benchmarks |
| Size | 16k Questions |
| License | MIT |
| Downloads | 12M |
| Tags | Benchmark, Multiple-Choice, 57-Subjects, STEM, Reasoning |
from datasets import load_dataset
ds = load_dataset("cais/mmlu")
Related datasets
- MATH (Hendrycks) — Benchmarking competition math ability
- Winogrande — Benchmarking commonsense reasoning
- HumanEval — Benchmarking Python code generation - never train on it
- MT-Bench — Judging multi-turn chat quality with LLM-as-judge
Frequently asked questions
Can I use MMLU (Massive Multitask Language Understanding) commercially?
Yes — MMLU (Massive Multitask Language Understanding) is released under MIT, a permissive license that allows commercial use, including training models you ship in a product. Check the dataset card for attribution requirements before release.
How much data does MMLU (Massive Multitask Language Understanding) contain, and do I need all of it?
MMLU (Massive Multitask Language Understanding) contains 16k Questions. It is an evaluation benchmark, so it is used in full to measure models — never mix it into training data, or your benchmark scores become meaningless.
What is MMLU (Massive Multitask Language Understanding) best used for?
Benchmarking general knowledge - never train on it. It belongs to the Evaluation & Benchmarks section of our dataset hub, where you'll find alternatives and complementary sets.
← All datasets | Fine-Tuning Guide