GSM8K (Grade School Math 8K) is a dataset of 8.5K high quality linguistically diverse grade school math word problems. The dataset was created to support the task of question answering on basic mathematical problems that require multi-step reasoning.
Dataset Card
By OpenAI, published under mit, revision 740312add88f.
Dataset Card for GSM8K
Table of Contents
- Dataset Description
- Dataset Summary
- Supported Tasks
- Languages
- Dataset Structure
- Data Instances
- Data Fields
- Data Splits
- Dataset Creation
- Curation Rationale
- Source Data
- Annotations
- Personal and Sensitive Information
- Considerations for Using the Data
- Social Impact of Dataset
- Discussion of Biases
- Other Known Limitations
- Additional Information
- Dataset Curators
- Licensing Information
- Citation Information
Dataset Description
- Homepage: https://openai.com/blog/grade-school-math/
- Repository: https://github.com/openai/grade-school-math
- Paper: https://arxiv.org/abs/2110.14168
- Leaderboard: [Needs More Information]
- Point of Contact: [Needs More Information]
Dataset Summary
Structure
main 8,792 rows
| Split | Rows | Size |
|---|---|---|
| train | 7,473 | 4.0 MB |
| test | 1,319 | 719.3 KB |
socratic 8,792 rows
| Split | Rows | Size |
|---|---|---|
| train | 7,473 | 5.2 MB |
| test | 1,319 | 943.2 KB |
Details
- Repository
- openai/gsm8k
- Publisher
- OpenAI
- Task category
- Text generation
- Tags
- math-word-problems
- Size category
- 1K<n<10K
- Languages
- en
- Revision
- 740312add88f781978c0658806c59bc2815b9866
- Last updated
- 2026-03-23
SAVRN's Notes on gsm8k
GSM8K is tiny, 17,584 rows and 5.9 MB in 7 files, and one of the most reused math datasets we track: 1.2M monthly downloads and a dozen inbound models, from SharperSwarm and CasualSwarms to a full Sheldon RLVR and SFT family built on Qwen2.5-3B-Instruct. There are two configs, main and socratic, each 8,792 rows with 7,473 train and 1,319 test, under an MIT license.
Its real role now is reward signal for RLVR rather than supervised SFT: grading is cheap and the answers are unambiguous. MIT license, no gate, and the revision is recent, last updated 2026-03-23, so diff against any copy you pinned years ago. Contamination is the tax: any base model plausibly saw GSM8K in pretraining, so treat absolute scores as marketing and require held-out math evals for real decisions.
Files
7 files, 5.9 MB in total.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| main/test-00000-of-00001.parquet | Data | 419.1 KB | ee7b8da9e381 |
| main/train-00000-of-00001.parquet | Data | 2.3 MB | ea82612ea958 |
| socratic/test-00000-of-00001.parquet | Data | 487.0 KB | 477dba702820 |
| socratic/train-00000-of-00001.parquet | Data | 2.7 MB | 54eb5fd2105a |
| README.md | Documentation | 7.9 KB | — |
| eval.yaml | Other | 920 B | — |
| .gitattributes | Repository | 1.6 KB | — |
License and Download
- License
- mit
- Access
- No access gate
Released by OpenAI through its official repository on Hugging Face. Read the license.
Models Trained on This Dataset
- Trained on (disclosed)SharperSwarm
- Trained on (disclosed)CasualSwarms
- Trained on (disclosed)Al-Khwarizmi-3B-GGUF
- Trained on (disclosed)shingi-27b
- Trained on (disclosed)Al-Khwarizmi-3B
- Trained on (disclosed)Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1
- Trained on (disclosed)Qwen2.5-3B-Instruct-Sheldon-SFT-touchup-v4-LoRA
- Trained on (disclosed)Qwen2.5-3B-Instruct-Sheldon-SFT-touchup-v4
- Trained on (disclosed)Qwen2.5-3B-Instruct-Sheldon-RLAIF-grpo-v2-LoRA
- Trained on (disclosed)Qwen2.5-3B-Instruct-Sheldon-RLAIF-grpo-v2
- Trained on (disclosed)Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1-LoRA
- Trained on (disclosed)Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v2-fullft
- Trained on (disclosed)Meet8-0.6B
- Trained on (disclosed)LMLM_97M_1
- Trained on (disclosed)bobcat-flash-1.2
- Trained on (disclosed)Qwen3.5-0.8B-Japanese-SFT-v2.5