Dave122/arxiv-math-preview

arXiv Math Papers — Preview Sample

⚠️ This is a ~1/5 preview of a full ~10M-token deliverable for an arXiv math scraping bounty. The full pipeline is built, tested, and the production dataset is already assembled locally. Send me the go-ahead and I'll publish the full dataset within an hour.

What's in the preview

• 150 math papers (1/5 of the planned ~721)
• 1,658,624 tokens (tiktoken cl100k_base)
• Same schema as the full deliverable
• Same category-balanced sampling strategy (sub-fields, no cap >40%)

Schema (one row per paper)

Field	Type	Description
arxiv_id	string	e.g. 2604.06242
arxiv_id_versioned	string	e.g. 2604.06242v1
title	string
authors	list[string]
abstract	string	from arXiv API
published / updated	string	ISO 8601
primary_category	string	e.g. math.AG
categories	list[string]	all cross-listed cats
abs_url / pdf_url / src_url	string	canonical arXiv URLs
body_text	string	cleaned body, LaTeX math preserved
latex_source	string	raw .tex source
pdf_bytes	binary	original arXiv PDF
token_count	int64	tiktoken cl100k_base count
main_tex_filename	string	inside the source tarball

Category breakdown (preview)

Primary category	Papers	Tokens	Share
math.GR	7	121,111	7.3%
math.OC	7	117,069	7.1%
math.GT	7	115,533	7.0%
math-ph	6	91,126	5.5%
math.DS	7	87,525	5.3%
math.AC	7	86,796	5.2%
math.CA	6	84,311	5.1%
math.AP	6	80,833	4.9%
math.MG	7	79,280	4.8%
math.RT	5	77,395	4.7%
math.KT	5	63,442	3.8%
math.CT	4	60,815	3.7%
math.SG	3	58,458	3.5%
math.RA	4	58,236	3.5%
math.FA	3	48,304	2.9%
math.DG	7	45,882	2.8%
math.AG	7	43,410	2.6%
math.PR	7	43,223	2.6%
math.CV	2	41,972	2.5%
math.NA	2	38,885	2.3%
math.ST	6	37,822	2.3%
math.AT	7	36,056	2.2%
math.LO	7	32,001	1.9%
math.SP	3	28,691	1.7%
math.HO	3	25,681	1.5%
math.CO	7	22,587	1.4%
math.NT	7	21,442	1.3%
math.OA	1	10,738	0.6%

Cleaning methodology

body_text was produced from each paper's main .tex file:

1. Open the source tarball, locate the file with \documentclass + \begin{document}.
2. Inline \input{...} / \include{...} one level deep.
3. Cut to body (\begin{document} ... \end{document}).
4. Strip LaTeX comments (escape-aware).
5. Remove \bibliography{...}, \begin{thebibliography}...\end{thebibliography}, \bibliographystyle{...}.
6. Replace figure / table / wrapfigure envs with [FIGURE_OR_TABLE].
7. Replace \cite*, \ref, \eqref, \label, etc. with [REF].
8. Math notation is preserved — inline $...$, display math, equation, align, gather, etc. are kept intact.
9. Whitespace collapsed; raw .tex is also kept in latex_source.

What the full deliverable will look like

• ~10,000,000 tokens of clean body text (tiktoken cl100k_base)
• ~720 papers (math papers run long, so this hits the token target with fewer papers; the bounty allowed "or enough to hit ~10M tokens")
• 28 distinct math sub-categories, max share ~13% (well under 40%)
• All math.* or math-ph primary categories
• Same Parquet/README structure as this preview
• Full PDFs + LaTeX sources embedded
• Built 2026-04-13T10:31:48.999121Z, ready to publish on request

Pipeline

Built entirely from scratch with a reproducible Python pipeline:

1. fetch_metadata.py — arXiv API per category (3s rate limit)
2. download_sources.py — .tex tarballs, 4 parallel workers, 3.1s/req
3. extract_text.py — clean body extraction with LaTeX math preserved
4. curate.py — tiktoken count + category-balanced selection
5. build_dataset.py — Parquet + README dataset card
6. upload_hf.py — HuggingFace push

Sources used: https://export.arxiv.org/api/query, https://arxiv.org/e-print/<id>, https://arxiv.org/pdf/<id>. All within the arXiv robots policy (≤4 parallel connections, ≥3s between requests).

Loading the preview

from datasets import load_dataset
ds = load_dataset("Dave122/arxiv-math-preview")
print(ds["train"][0].keys())

License

arXiv non-exclusive license. Each paper retains its original authorship and license (https://arxiv.org/help/license).