$ open /benchmarks/math▌discussed░░░░░░░░░░░░░░░░░░░░ 0/5000 problems · discussed · posts
MATH
Mathematics
MATH organises 12,500 high-school competition problems into seven subjects and five difficulty levels, each with a full worked solution. It underlies much of the later mathematics evaluation work; this catalogue imports the full test split: 5,000 problems across all seven subjects.
Dan Hendrycks and collaborators introduced MATH in 2021 to test mathematical problem solving beyond routine word problems. They assembled competition mathematics with worked solutions so that evaluation could address harder reasoning and training could use intermediate steps. EleutherAI maintains the Hugging Face distribution recorded here, whose 2023 publication date is later than the original benchmark.
How it was built
MATH collects 12,500 high-school competition problems with worked solutions, organised into seven subjects and five difficulty levels. The material comes from existing competition problem collections rather than newly generated model outputs. The catalogue contains a 200-problem selection from the algebra test split, whose source contains 1,187 problems; that selection is not the full benchmark.
At a glance
Metric
accuracy on the final boxed answer
Size
5,000 problems imported from the test split across all seven subjects.
License
mit (declared on the Hugging Face dataset card)
Released
Dataset published on Hugging Face 2023-09
Maintainer
EleutherAI (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
61.6 EM — DeepSeek-V3 (vendor, self-reported)
What a problem looks like
A problem statement in LaTeX, with its subject and difficulty level. The worked solution is not part of the prompt.
The problems are written for students who have the technique and need to choose it. Difficulty is stated per problem, so level 5 items are hard by construction rather than by accident.
Limitations and criticism
Final-answer accuracy does not establish that the reasoning leading to an answer is valid. Answer extraction and equivalence checking can affect whether mathematically equivalent responses receive credit. Public problems and worked solutions create a risk of training exposure. Results on an algebra selection or another reduced subset should be distinguished from results on the full test set.
How much room is left
The recorded vendor report gives DeepSeek-V3 base 61.6 percent exact match with four examples in the prompt. This leaves considerable headroom for that setup, but one base-model report is insufficient to establish the ceiling for stronger reasoning systems. Comparisons need to identify the evaluated subset and whether the model received tools or multiple attempts.
@article{hendrycks2021math,
title={Measuring Mathematical Problem Solving With the MATH Dataset},
author={Dan Hendrycks and Collin Burns and Saurav Kadavath and Akul Arora and Steven Basart and Eric Tang and Dawn Song and Jacob Steinhardt},
journal={arXiv preprint arXiv:2103.03874},
year={2021}
}
Related benchmarks
Discussion
1 thread and 0 posts on this page. Anyone may reply; no account is required.
Contribute
Share a reproduction, report an evaluation issue, or suggest a correction to this benchmark overview.