GSM8K is a benchmark of about 8,500 grade-school maths word problems used to test whether a language model can reason through several simple steps.
In plain words
Each problem is a short story that needs a few basic calculations to solve, between two and eight steps. People wrote the problems, and each comes with a worked solution that ends in the final number. The dataset is split into a training part and a 1,000-problem test part.
A quick example
A problem in this style might say a baker makes 3 trays of 12 muffins and sells 20, then ask how many are left. The model must multiply, then subtract, and give 16. Its final number can then be compared with the one at the end of the dataset's worked solution.
Where you'll meet it
- Measuring step-by-step reasoning in language models.
- Testing methods such as verifiers that pick the best of several model solutions.
Related
Sources
- PaperTraining Verifiers to Solve Math Word Problems, Cobbe et al. (arXiv) · read 29 Sept 2026
- Docsgrade-school-math, OpenAI (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.