Approximately 860k math problems, where each solution is formatted in a Chain of Thought (CoT) manner. The sources of the dataset range from Chinese high school math exercises to US and international mathematics olympiad competition problems.
Dataset Card
By Project-Numina, published under apache-2.0, revision 9d8d210c9f6a.
Dataset Card for NuminaMath CoT
Dataset Description
- Homepage: https://projectnumina.ai
- Repository: https://github.com/project-numina/aimo-progress-prize
- Paper: https://github.com/project-numina/aimo-progress-prize/blob/main/report/numina_dataset.pdf
- Leaderboard:
- Point of Contact: Jia Li
Dataset Summary
Approximately 860k math problems, where each solution is formatted in a Chain of Thought (CoT) manner. The sources of the dataset range from Chinese high school math exercises to US and international mathematics olympiad competition problems. The data were primarily collected from online exam paper PDFs and mathematics discussion forums. The processing steps include (a) OCR from the original PDFs, (b) segmentation into problem-solution pairs, (c) Translation into English, (d) realignment to produce a CoT reasoning format, and (e) final answer formatting.
Source breakdown
| Source | Number of Samples |
|---|---|
| aops_forum | 30201 |
| amc_aime | 4072 |
| cn_k12 | 276591 |
| gsm8k | 7345 |
| math | 7478 |
| olympiads | 150581 |
| orca_math | 153334 |
| synthetic_amc | 62111 |
| synthetic_math | 167895 |
| Total | 859608 |
Licensing Information
The dataset is available under the Apache License, Version 2.0.
Citation Information
Structure
default 859,594 rows
| Split | Rows | Size |
|---|---|---|
| train | 859,494 | 2.5 GB |
| test | 100 | 282.4 KB |
Details
- Repository
- AI-MO/NuminaMath-CoT
- Publisher
- Project-Numina
- Task category
- Text generation
- Tags
- aimo, math
- Size category
- Not stated by the source
- Languages
- en
- Revision
- 9d8d210c9f6a36c8f3cd84045668c9b7800ef517
- Last updated
- 2024-11-25
SAVRN's Notes on NuminaMath-CoT
NuminaMath-CoT is supervised fine-tuning fuel: 859,594 math problems with chain-of-thought solutions, 1.2 GB, 8 files, 187.3k monthly downloads, tagged aimo and math, with text generation as the stated task. The source breakdown on the card tells you the flavor, dominated by cn_k12 at 276,591 samples, then aops_forum at 30,201 and amc_aime at 4,072, plus gsm8k, with the card's pipeline running OCR from the original exam PDFs, translation to English and realignment into CoT format.
Apache 2.0 makes it commercially clean, and 1.2 GB is cheap to store and ship. The disclosed inbound models show the intended path: slm-125m-base, slm-125m-sft and slm-125m-dpo are one lineage moving from base through SFT to preference tuning on this data. This is reasoning-format SFT data, not pretraining volume; plan one fine-tuning pass, not a corpus to mine. English only, last updated 2024-11-25.
Files
8 files, 1.2 GB in total.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| data/test-00000-of-00001.parquet | Data | 165.9 KB | 62f94a9529fa |
| data/train-00000-of-00005.parquet | Data | 246.6 MB | 25c0fa2ca96d |
| data/train-00001-of-00005.parquet | Data | 246.9 MB | 2d4ba17018b6 |
| data/train-00002-of-00005.parquet | Data | 246.8 MB | c607c1837954 |
| data/train-00003-of-00005.parquet | Data | 246.9 MB | e07293604d1a |
| data/train-00004-of-00005.parquet | Data | 246.9 MB | e3929523d643 |
| README.md | Documentation | 2.7 KB | — |
| .gitattributes | Repository | 2.3 KB | — |
License and Download
- License
- apache-2.0
- Access
- No access gate
Released by Project-Numina through its official repository on Hugging Face. Read the license.
Models Trained on This Dataset
- Trained on (disclosed)slm-125m-dpo
- Trained on (disclosed)slm-125m-sft
- Trained on (disclosed)slm-125m-base