| --- |
| license: odc-by |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - math |
| - code |
| - pre-training |
| - synthesis |
| size_categories: |
| - 1B<n<10B |
| --- |
| |
| # MegaMath: Pushing the Limits of Open Math Copora |
|
|
| > Megamath is part of TxT360, curated by LLM360 Team. |
| <center><img src="teasor.png" alt="MegaMath Collection" /></center> |
|
|
| We introduce MegaMath, an open math pretraining dataset curated from diverse, math-focused sources, with over 300B tokens. |
| MegaMath is curated via the following three efforts: |
| - **Revisiting web data**: |
| We re-extracted mathematical documents from Common Crawl with math-oriented HTML optimizations, fasttext-based filtering and deduplication, all for acquiring higher-quality data on the Internet. |
| - **Recalling Math-related code data**: |
| We identified high quality math-related code from large code training corpus, Stack-V2, further enhancing data diversity. |
| - **Exploring Synthetic data**: |
| We synthesized QA-style text, math-related code, and interleaved text-code blocks from web data or code data. |
|
|
| ## MegaMath Compared to Existing Datasets |
| MegaMath is the largest open math pre-training dataset to date, surpassing DeepSeekMath (120B) tokens. |
| <div style="display: flex; justify-content: center; gap: 20px;"> |
| <img src="https://cdn-uploads.huggingface.co/production/uploads/628f6e5ab90dde28ef57d293/lFa_r4gSXhjwep7XAwwQj.png" width="75%" /> |
| </div> |
|
|
| ## MegaMath Delivers with High Quality |
| During development, we use extensive experiments to find optimal practice for text extraction, deduplication, fasttext training, etc. Training MegaMath data shows better performance than existing open datasets. |
|
|
| <div style="display: flex; justify-content: center; gap: 20px;"> |
| <img src="https://cdn-uploads.huggingface.co/production/uploads/628f6e5ab90dde28ef57d293/-E1tZP-vbU1ZPzy56cl4s.png" width="30%" /> |
| <img src="https://cdn-uploads.huggingface.co/production/uploads/628f6e5ab90dde28ef57d293/XSBJ_wVexM-0rk9bcpU5Q.png" width="30%" /> |
| </div> |
|
|
|
|
| ## Training MegaMath on Latest LMs |
| We also release two proof-of-concept models which is based on [Llama-3.2-1B](https://huggingface.co/meta-llama/Llama-3.2-1B) and [LLama-3.2-3B](https://huggingface.co/meta-llama/Llama-3.2-3B). |
| Training MegaMath on Llama-3.2-1B and LLama-3.2-3B brings about 15% ~ 20% performance boost on 10 downstream benchmarks, demonstrateing its high data quality. |
| <div style="display: flex; justify-content: center; gap: 20px;"> |
| <img src="https://cdn-uploads.huggingface.co/production/uploads/628f6e5ab90dde28ef57d293/EIReQ8TIbyn1V3JfsEKiL.png" width="50%" /> |
| </div> |
|
|
| ## Detailed Statistics |
| | **Category** | **# Sample(M)** | **# Toks(B)** | **Avg. (# Toks)** | |
| |------------------------|----------------:|--------------:|------------------:| |
| | **Web Domain** | **121.5** | **279.0** | **2296.9** | |
| | Web | 106.5 | 263.9 | 2478.7 | |
| | Web-Pro | 15.0 | 15.1 | 1006.0 | |
| | **Code Domain** | **13.4** | **28.1** | **2102.7** | |
| | **Synthetic Data** | **80.2** | **64.5** | **804.5** | |
| | Translated Code | 7.4 | 7.2 | 979.5 | |
| | Q&A | 22.6 | 7.0 | 308.3 | |
| | Text&Code Block | 50.2 | 50.3 | 1002.1 | |
| | **Total** | **215.1** | **371.6** | **1727.6** | |
|
|
|
|
| ## Citation |
| If you use our dataset or find our work useful, please cite |
| ```bibtex |
| @article{zhou2025megamath, |
| title = {MegaMath: Pushing the Limits of Open Math Corpora}, |
| author = {Zhou, Fan and Wang, Zengzhi and Ranjan, Nikhil and Cheng, Zhoujun and Tang, Liping and He, Guowei and Liu, Zhengzhong and Xing, Eric P.}, |
| journal = {arXiv preprint arXiv:2504.02807}, |
| year = {2025}, |
| note = {Preprint} |
| } |
| ``` |
|
|
