slabhead/math-nt-papers

arXiv Number Theory Papers

A focused dataset of recent math.NT (number theory) papers from arXiv, prepared for language-model training and research. Companion to math-papers-10m, but narrowed to a single subfield for dense topical coverage.

Summary

• Category: math.NT (primary)
• Total papers: 500
• Total body tokens (cl100k_base): 16,400,635
• Source: arXiv (https://arxiv.org), via the arxiv Python client + content endpoints (/html/, /e-print/, /pdf/)
• Generated: 2026-04-11
• License: Papers are redistributed under arXiv's non-exclusive license to distribute. See https://arxiv.org/help/license. Cite the original authors.

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
title	string	Paper title
authors	list[string]	Author names
abstract	string	Paper abstract
primary_category	string	Always math.NT (primary category)
categories	list[string]	All arXiv subject tags (may include cross-listings)
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 as LaTeX in $...$)
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 tarball is attached
has_pdf	bool	Whether original PDF is attached
tex_bytes_len	int	Size of .tex tarball on the repo
pdf_bytes_len	int	Size of PDF on the repo

Raw source artifacts live alongside the Parquet data:

sources/tex/<arxiv_id>.tar.gz
sources/pdf/<arxiv_id>.pdf

Primary-category breakdown

Category	Papers	Share
math.NT	361	72.2%
math.AG	42	8.4%
math.CO	34	6.8%
math.RT	13	2.6%
math.DS	12	2.4%
math.LO	5	1.0%
cs.IT	3	0.6%
math.SP	3	0.6%
math.GT	3	0.6%
math.PR	3	0.6%
hep-th	2	0.4%
math.AC	2	0.4%
cs.CR	2	0.4%
math.KT	2	0.4%
math-ph	2	0.4%
math.CA	2	0.4%
cs.MS	2	0.4%
math.HO	1	0.2%
math.CV	1	0.2%
cs.FL	1	0.2%
quant-ph	1	0.2%
math.QA	1	0.2%
math.MG	1	0.2%
math.FA	1	0.2%

Year distribution (submission year)

Year	Papers
2026	500

Methodology

1. Metadata: Fetched via the arxiv Python client (ToS-compliant 3.1s delay, retries on 429/503). Query: cat:math.NT, sorted by submission date descending.
2. Body extraction cascade (per paper):
   • HTML-first: arxiv.org/html/<id> is fetched and parsed with BeautifulSoup. MathML <math> elements are replaced with their LaTeX source (from the <annotation encoding="application/x-tex"> child) so each equation appears once, not duplicated as both Unicode rendering and LaTeX source.
   • pylatexenc fallback: If HTML is unavailable, the /e-print/ tarball is unpacked and the main .tex is converted via pylatexenc with math preserved in delimited form.
   • pymupdf fallback: Last resort — body text extracted from the PDF via PyMuPDF (math renders as plaintext garbage but guarantees coverage).
3. Body filter: Reject bodies below 1000 chars or 300 tokens.
4. Token count: tiktoken.get_encoding("cl100k_base").
5. Artifacts: Every paper gets its raw .tar.gz e-print and original .pdf uploaded alongside the Parquet row.

Extraction method breakdown

• html: 450
• pylatexenc: 46
• pymupdf: 4

Caveats

• Bodies retain references to figures/tables; figures themselves are not extracted.
• Bibliographies are stripped pre-extraction.
• Math coverage depends on extractor: HTML produces the cleanest output because it uses the LaTeXML-generated LaTeX source directly; pylatexenc is a close second; pymupdf (PDF fallback) produces the weakest math.
• Date range reflects the most recent papers on arXiv at the time of scraping (see "Year distribution" above).

Provenance

• Generated by scripts/08_nt_dataset.py in the math-papers-pipeline repo.
• Reproducible: same arXiv query + fixed sort order + fixed batch size.