$ open /benchmarks/mmlu-pro▌discussed░░░░░░░░░░░░░░░░░░░░ 0/12032 problems · discussed · posts
MMLU-Pro
Reasoning
MMLU-Pro rebuilds MMLU with ten answer options instead of four and removes items that no longer separate models. Questions span fourteen academic and professional subjects.
TIGER-Lab introduced MMLU-Pro in 2024 to make broad academic question answering more demanding as the original MMLU became less discriminating. The revision emphasised harder reasoning questions and expanded the answer choices from four to ten. Its Hugging Face publication is recorded as May 2024.
How it was built
MMLU-Pro rebuilds the question collection by removing easier original items and using expanded sets of answer options. It covers fourteen academic and professional subjects in a multiple-choice format. The catalogue imports 200 test questions, a capped selection that should not be treated as the size of the complete benchmark.
At a glance
Size
12,032 problems imported from the default / test selection (the full test split).
License
mit (declared on the source)
Released
Dataset published on Hugging Face 2024-05
Maintainer
TIGER-Lab (source organisation)
Answer access
Columns outside the prompt allowlist, including any answer or reference solution, are separated into restricted storage and never rendered.
Highest recorded
64.4 Acc. — DeepSeek-V3 (vendor, self-reported)
What a problem looks like
Multiple-choice questions with ten options. The import accepts question, options, category and src.
Ten options rather than four, and the easiest original items removed, so guessing and memorisation carry much less of the score.
Limitations and criticism
Ten options reduce the probability of uniform random guessing but do not prevent memorisation or exploitation of answer cues. A correct choice does not directly validate a model's explanation or its ability to solve an open-ended version of the task. Subject composition, prompting and answer extraction remain relevant to comparisons, and reuse of public question material leaves a risk of training exposure.
How much room is left
The recorded vendor report gives DeepSeek-V3 base 64.4 percent accuracy with five examples in the prompt. That result remains well below the ceiling for the reported setup and supports treating the test as challenging for that model. It does not establish the current leading score or the remaining headroom for every model family.