slabhead/math-papers-10m

Math Papers 10M

A curated dataset of recent mathematics papers from arXiv, prepared for language-model training and research. Target: ~10M cl100k_base tokens of clean body text across a balanced set of math subcategories.

Summary

• Total papers: 901
• Total body tokens (cl100k_base): 33,121,411
• Source: arXiv (https://arxiv.org), via the public metadata API and content endpoints (/html/, /e-print/, /pdf/)
• Generated: 2026-04-12
• License: Papers are redistributed under arXiv's non-exclusive license to distribute. See https://arxiv.org/help/license. Cite the original authors when using this data.

Contents

Each batch is a Parquet file under data/:

data/batch-0001.parquet
data/batch-0002.parquet
...

Per-row schema:

field	type	description
arxiv_id	string	Canonical arXiv identifier (e.g. 2401.12345)
title	string	Paper title
authors	list[string]	Author names
abstract	string	Paper abstract
primary_category	string	arXiv primary subject (e.g. math.AG)
categories	list[string]	All arXiv subject tags
published	string	ISO-8601 submission timestamp
updated	string	ISO-8601 last-updated timestamp
abs_url	string	arXiv abstract URL
pdf_url	string	arXiv PDF URL
body	string	Cleaned body text (math preserved where possible)
body_chars	int	Character length of body
body_tokens	int	Token count via tiktoken cl100k_base
extraction_method	string	html, pylatexenc, or pymupdf
has_tex	bool	Whether original .tex source was retrieved
has_pdf	bool	Whether original PDF was retrieved

Original artifacts (when available) are stored alongside the data:

sources/tex/<arxiv_id>.tar.gz   # original .tex tarball from arXiv
sources/pdf/<arxiv_id>.pdf      # original rendered PDF

Category Breakdown

Primary category	Papers	%
math.AG	61	6.8%
math.CO	51	5.7%
math.LO	47	5.2%
math.NT	46	5.1%
math.RT	45	5.0%
math.GT	42	4.7%
math.PR	41	4.6%
math.NA	41	4.6%
math.DG	37	4.1%
math.OC	36	4.0%
math.FA	32	3.6%
math.CA	31	3.4%
math.DS	30	3.3%
math.AT	30	3.3%
math.RA	30	3.3%
math.CT	27	3.0%
math-ph	27	3.0%
math.QA	25	2.8%
math.ST	24	2.7%
math.SG	23	2.6%
math.AP	19	2.1%
math.GR	14	1.6%
cs.LG	13	1.4%
hep-th	12	1.3%
stat.ML	10	1.1%
quant-ph	9	1.0%
eess.SY	8	0.9%
math.OA	7	0.8%
cs.LO	6	0.7%
cond-mat.stat-mech	6	0.7%
cs.IT	6	0.7%
math.AC	5	0.6%
cs.DS	5	0.6%
math.KT	5	0.6%
stat.ME	4	0.4%
math.SP	4	0.4%
cs.CC	4	0.4%
gr-qc	3	0.3%
math.CV	3	0.3%
physics.ao-ph	3	0.3%
cond-mat.soft	2	0.2%
cs.AI	2	0.2%
math.MG	2	0.2%
math.HO	2	0.2%
cs.CV	2	0.2%
cs.RO	2	0.2%
nlin.AO	1	0.1%
astro-ph.IM	1	0.1%
cs.DB	1	0.1%
physics.atom-ph	1	0.1%
stat.CO	1	0.1%
physics.plasm-ph	1	0.1%
physics.data-an	1	0.1%
nlin.PS	1	0.1%
cs.GT	1	0.1%
q-fin.PM	1	0.1%
astro-ph.CO	1	0.1%
nucl-th	1	0.1%
astro-ph.EP	1	0.1%
cs.CG	1	0.1%
cs.ET	1	0.1%
physics.flu-dyn	1	0.1%
cond-mat.other	1	0.1%

No single subcategory exceeds 40% of the total.

Methodology

1. Metadata. Queried export.arxiv.org/api/query for each of ~20 math subcategories (including math-ph), sorted by submission date descending, pulling up to 150 papers per category. Recent papers are preferred.
2. Balanced sampling. Papers are round-robin sampled across categories with a 40% cap on any single category.
3. Content fetch. For each sampled paper we try, in order:
   • arxiv.org/html/<id> — native HTML5 render with math preserved
   • arxiv.org/e-print/<id> — .tex source tarball
   • arxiv.org/pdf/<id>.pdf — rendered PDF
4. Extraction cascade.
   • HTML path: BeautifulSoup + lxml, stripping nav, bibliography, author metadata, figures, and page headers/footers.
   • TeX path: pylatexenc with math_mode="with-delimiters" after cutting off \bibliography / thebibliography sections.
   • PDF path: pymupdf, skipping the title page and the last two pages (references / appendices).
5. Quality gate. Bodies must have at least 1000 characters and 300 cl100k_base tokens. Papers failing the cascade at every level are logged and excluded.
6. Tokenization. All token counts use tiktoken with the cl100k_base encoding (the encoding used by GPT-4 / GPT-3.5-turbo).
7. Rate limiting. All arXiv requests respect a 3-second minimum delay with exponential backoff on 429 / 503 responses.

Known Limitations

• pymupdf-based extractions lose math structure — equations become flat text. The extraction_method field lets you filter these out if desired.
• Figures, tables, and captions are dropped from the body text in all extraction paths.
• Some older papers do not have HTML or TeX sources; these rely on the PDF fallback.
• The dataset is a snapshot — arXiv IDs and categories can be updated later by authors. Re-run the pipeline to refresh.

Intended Use

Research, pretraining, and fine-tuning of models that handle mathematical text. Not a substitute for reading the original papers.

Citation

Please cite the original arXiv papers when using individual rows. If you need to cite this specific extraction, use:

@dataset{math_papers_10m_20260412,
  title   = {Math Papers 10M},
  year    = 2026,
  url     = {https://huggingface.co/datasets/slabhead/math-papers-10m}
}