AutoMathText: 200 GB of Mathematical Text from Web, arXiv, and GitHub

AutoMathText is a dataset of approximately 200 GB of mathematical texts compiled from sources including various websites, arXiv, and GitHub repositories like OpenWebMath, RedPajama, and Algebraic Stack. The dataset was created by author 'math-ai' and its associated work was accepted to ACL 2025 Findings. The dataset was last updated on July 16, 2025.

Use Cases

• Pretraining language models based on the extensive collection of mathematical text.
• Fine-tuning models for mathematical question answering based on the described textual content.
• Conducting research on mathematical reasoning based on the curated compilation from diverse platforms.

Strengths

• Approximately 200 GB of data, indicating a substantial scale.
• Curated from a diverse range of sources including arXiv, GitHub, and various websites.
• Associated research was accepted to a major conference (ACL 2025 Findings).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and specific license details are unknown.
• Data may reflect source bias inherent to the compilation from web, arXiv, and GitHub.

Provenance

Source

Compilation from various websites, arXiv, and GitHub repositories (OpenWebMath, RedPajama, Algebraic Stack).

Collection Method

Curated using the AutoDS method described in the associated ACL 2025 Findings paper.

Freshness

Last updated 2025-07-16 05:03:38.