Keiran Paster*, Marco Dos Santos*, Zhangir Azerbayev, Jimmy Ba
GitHub | ArXiv
| PDF
OpenWebMath is a dataset containing the majority of the high-quality, mathematical text from the internet. It is filtered and extracted from over 200B HTML files on Common Crawl down to a set of 6.3 million documents containing a total of 14.7B tokens. OpenWebMath is intended for use in pretraining and finetuning large language models.
You can download the dataset using Hugging Face:
from datasets import… See the full description on the dataset page:
https://huggingface.co/datasets/Alignment-Lab-AI/Open-Web-Math.
