MMLU-Pro
MMLU-Pro is an evaluation benchmark built by TIGER-Lab and released in 2024 as a more demanding successor to the original MMLU. It contains 12,032 multiple-choice questions spanning 14 disciplines, including mathematics, physics, chemistry, law, engineering, economics, health, psychology, business, biology, philosophy, computer science and history. Two changes make it harder than its predecessor: it raises the number of answer options from four to up to ten, which lowers the odds of guessing correctly, and it leans towards reasoning-heavy problems drawn from the original MMLU, STEM websites, TheoremQA and SciBench, with unreasonable or noisy options pruned during review.
Access is straightforward and fully open. The dataset lives on Hugging Face under TIGER-Lab/MMLU-Pro, auto-converted to Parquet and readable through the datasets, pandas or Polars libraries. There is no account wall, application or competition gate, so it counts as open access.
For RAG, MMLU-Pro is most useful as a retrieval-augmented evaluation set: you can measure whether adding a retriever improves accuracy on broad, knowledge-intensive questions, and probe where retrieval helps versus where pure reasoning is needed. For an AI tutor it is an assessment and diagnostic resource rather than teaching material. The 14 disciplines give a broad, multi-subject bank of practice items for benchmarking a tutor's subject coverage or for mastery modelling across topics, but the questions are designed to test models, not to sequence a curriculum or grade content by level, so it complements instructional sources such as openstax, khan-academy, ck-12 and siyavula rather than replacing them.
The licence is MIT, so commercial use is permitted provided you retain the licence and copyright notice; there is no share-alike or non-commercial restriction. Note that questions are aggregated from several upstream sources, so check provenance if you need strict rights clarity. Alongside our maths sources (amps, stackmathqa, megamath, naturalproofs), MMLU-Pro is the broad, cross-subject yardstick.
Related sources
CRAG (Comprehensive RAG Benchmark)
Meta's factual question answering benchmark of 4,409 QA pairs across five domains, paired with mock web-search and knowledge-graph retrieval APIs. It is distinct from "Corrective RAG (CRAG)", which is a retrieval method, not a dataset.
FinanceBench
A benchmark for evaluating LLMs and retrieval systems on questions about public financial filings. The full set holds 10,231 question, answer and evidence triplets; a 150 example sample is released openly.
FlashRAG Benchmark Datasets
A research toolkit that bundles standardised, pre-processed benchmark datasets for RAG evaluation across many scenarios. Every dataset comes in a single unified JSONL format, so you can swap benchmarks in and out without rewriting your data loading each time.
FRAMES (Fact, Fetch, and Reason)
A Google evaluation set of 824 challenging questions that tests factuality, multi-hop retrieval and reasoning together, each paired with a gold answer and the Wikipedia articles needed to answer it. Built to measure end-to-end RAG.