Details

Population & demographics

Sample size

8,500 problems

Age range

5 - 12 years old

Grades 1 - 6

Languages

English

Metadata per student

What demographic metadata is available on an individual level?

Available

None

Unavailable

Age, Ethnicity, Gender, Geography, Learning disability, Native language, Race, Socioeconomic status, Speech pathology

Data collection & preparation

School subjects

Math

Collection protocol

Problems were written by human problem writers.

Opt-out was available

Unknown or not applicable

Data processing

Questions underwent extensive quality control and were selected to curate a high diversity of questions with moderate difficulty.

Tips for using the data

Technical tips

Each answer contains the mutli-step solution to the problem. The final answer is the last line of the answer text, preceded by ####. Calculations are represented using calculator notation (<<calculation>>). To remove the calculator annotation, remove every string surrounded by << >>.

For additional usage information refer to GitHub and HuggingFace.

Appropriate uses and limitations

This dataset is useful for assessing math reasoning of LLM modeling approaches.

Helpful links

Paper

GitHub

Hugging Face

Authorship

Organization

OpenAI

Authors

Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, John Schulman, Karl Cobbe

Release date

Nov. 18, 2021

Citation

Copy BibTex