Convex Markets / Datasets / Reasoning QA / MATH (Hendrycks)
SKU RQA-5002 · Sold by External

MATH (Hendrycks)

Product specifications

SKURQA-5002
Data typeReasoning QA
Volume12,500 problems (7,500 train / 5,000 test)
Size on disk~4.9 MB (Parquet download across all 7 configs); ~9.7 MB uncompressed (per HF card byte counts)
FormatParquet (Hugging Face mirror, 7 subject configs); original repository ships per-problem JSON files with fields problem/level/type/solution
Access modelPUBLIC LICENSE
PricingFree · open dataset (HF)
Quality score
LicenseMIT License
MATH is a benchmark of 12,500 problems drawn from high-school mathematics competitions (e.g. AMC 10/12, AIME), released with the paper 'Measuring Mathematical Problem Solving With the MATH Dataset' (Hendrycks et al., NeurIPS 2021). Each problem is labeled with one of 7 subjects (algebra, counting_and_probability, geometry, intermediate_algebra, number_theory, prealgebra, precalculus) and a difficulty from 'Level 1' to 'Level 5'. Every problem ships with a complete natural-language, LaTeX-formatted worked solution ending in the gold final answer enclosed in \boxed{...}. The dataset is split into 7,500 training and 5,000 test problems. This catalog listing is built from the EleutherAI/hendrycks_math Hugging Face mirror (Parquet, one config per subject) which carries the MIT license and the original https://github.com/hendrycks/math release. Models are typically scored by extracting the boxed final answer and checking mathematical equivalence against the gold answer.