GSM8K
OpenAIGrade School Math 8K: 8,500 grade-school math word problems, 7,473 for training and 1,319 for test, used to evaluate chain-of-thought reasoning on concise single-answer tasks. It ships main and socratic configurations and comes from the paper on training verifiers to solve math word problems.
Openness
5 high confidence- license
- MIT(OSI-style permissive, fully redistributable)
- redistributability
- full(downloadable parquet, ungated)
- datasheet
- yes(dataset card)
- size
- 8.5K problems
- splits
- train/test public
Fully open: MIT, ungated, downloadable, with a dataset card. Clean open-data case.
- https://huggingface.co/datasets/openai/gsm8k recorded 2026-09-29
The dataset page carries the tag "license:mit", its Licensing Information section reads "The GSM8K dataset is licensed under the", and its card opens "GSM8K (Grade School Math 8K) is a dataset of 8.5K high quality linguistically diverse grade school math word problems".
Adoption
5 high confidenceAdoption is measured through Hugging Face downloads of openai/gsm8k, though this category counts dataset downloads on a far smaller scale than it does for software or models. GSM8K remains the standard grade-school math-reasoning benchmark, cited in nearly every LLM release report.
- https://huggingface.co/api/datasets/openai/gsm8k recorded 2026-09-01
1,167,043 downloads in the trailing 30 days for openai/gsm8k
Capability
not assessedA dataset is not 'capable', so capability is left unscored. GSM8K is now largely saturated by frontier models, but that is a quality caveat rather than a capability score.
- https://huggingface.co/datasets/openai/gsm8k recorded 2026-09-29
The dataset card states "GSM8K (Grade School Math 8K) is a dataset of 8.5K high quality linguistically diverse grade school math word problems"; the page embeds a Hub leaderboard ranking models on the dataset and says "It has been used for many benchmarks, including the", but states no accuracy, throughput or feature claim for the corpus itself.
Verified 2026-09-27