MATH dataset from https://github.com/hendrycks/math
Dataset Card
By EleutherAI, published under mit, revision 21a5633873b6.
Dataset Summary
MATH dataset from https://github.com/hendrycks/math
Citation Information
@article{hendrycksmath2021,
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={NeurIPS},
year={2021}
}
Structure
algebra 2,931 rows
| Split | Rows | Size |
|---|---|---|
| train | 1,744 | 966.5 KB |
| test | 1,187 | 649.4 KB |
counting_and_probability 1,245 rows
| Split | Rows | Size |
|---|---|---|
| train | 771 | 667.4 KB |
| test | 474 | 353.8 KB |
geometry 1,349 rows
| Split | Rows | Size |
|---|---|---|
| train | 870 | 1.1 MB |
| test | 479 | 523.1 KB |
intermediate_algebra 2,198 rows
| Split | Rows | Size |
|---|---|---|
| train | 1,295 | 1.2 MB |
| test | 903 | 795.1 KB |
number_theory 1,409 rows
| Split | Rows | Size |
|---|---|---|
| train | 869 | 595.8 KB |
| test | 540 | 349.5 KB |
prealgebra 2,076 rows
| Split | Rows | Size |
|---|---|---|
| train | 1,205 | 723.8 KB |
| test | 871 | 510.2 KB |
precalculus 1,292 rows
| Split | Rows | Size |
|---|---|---|
| train | 746 | 816.2 KB |
| test | 546 | 552.9 KB |
Details
- Repository
- EleutherAI/hendrycks_math
- Publisher
- EleutherAI
- Task category
- Not stated by the source
- Tags
- Not stated by the source
- Size category
- Not stated by the source
- Languages
- Not stated by the source
- Revision
- 21a5633873b6a120296cce3e2df9d5550074f4a3
- Last updated
- 2025-01-12
SAVRN's Notes on hendrycks_math
EleutherAI's repackaging of the Hendrycks MATH benchmark is 12,500 problems, 4.9 MB, MIT licensed, organized as seven subject configs — algebra (2,931 rows), counting_and_probability, geometry, intermediate_algebra, number_theory, prealgebra and precalculus — each with its own train/test cut. Algebra alone splits 1,744 train / 1,187 test. The hub card is terse, pointing at the original math repository, but the structure table does the work.
Use the train halves for math SFT and keep the test halves sacred — they are the de facto frontier-model math benchmark, and training on them poisons exactly the number people compare. The per-subject layout lets you ablate single subjects cheaply instead of running the full 12,500. At 4.9 MB and 231.2k monthly downloads this is pure evaluation infrastructure; storage and bandwidth are non-issues.
Files
16 files, 4.9 MB in total.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| algebra/test-00000-of-00001.parquet | Data | 353.1 KB | 0e04a850dac9 |
| algebra/train-00000-of-00001.parquet | Data | 505.2 KB | 73347f5b5586 |
| counting_and_probability/test-00000-of-00001.parquet | Data | 175.0 KB | c324818949c7 |
| counting_and_probability/train-00000-of-00001.parquet | Data | 329.4 KB | d37ae017930d |
| geometry/test-00000-of-00001.parquet | Data | 263.8 KB | 98720f3a8134 |
| geometry/train-00000-of-00001.parquet | Data | 549.4 KB | 9dbf5e5b5549 |
| intermediate_algebra/test-00000-of-00001.parquet | Data | 394.8 KB | e859f72486c9 |
| intermediate_algebra/train-00000-of-00001.parquet | Data | 575.2 KB | 0c87f3798ebd |
| number_theory/test-00000-of-00001.parquet | Data | 182.0 KB | ba156c9c9fa2 |
| number_theory/train-00000-of-00001.parquet | Data | 308.7 KB | 81c08649149b |
| prealgebra/test-00000-of-00001.parquet | Data | 267.7 KB | 301e5d8ade5c |
| prealgebra/train-00000-of-00001.parquet | Data | 383.6 KB | 765b526fb5c0 |
| precalculus/test-00000-of-00001.parquet | Data | 242.2 KB | da37474ba376 |
| precalculus/train-00000-of-00001.parquet | Data | 353.7 KB | 28bb1ff95525 |
| README.md | Documentation | 3.9 KB | — |
| .gitattributes | Repository | 2.3 KB | — |
License and Download
- License
- mit
- Access
- No access gate
Released by EleutherAI through its official repository on Hugging Face. Read the license.