ragarwal/arxiv-markdown

arXiv Markdown

Last updated: 22 September 2026

Full text of 1,730,536 arXiv papers as Markdown, one row per paper with its metadata. The subset covers 58 categories in computer science, electrical engineering, mathematics, statistics, physics, condensed matter, quantum physics, astrophysics instrumentation and quantitative biology, from 1991 to September 2026.

Papers with LaTeX source were converted with latexml and pandoc, so formulas are LaTeX in $…$ and $$…$$. Papers that exist on arXiv only as PDF, about 12 %, were transcribed with the Chandra OCR 2 model; figures in those are short text descriptions.

Loading

The full dataset is about 28 GB of parquet. Hugging Face datasets converts it into a local Arrow cache of about 104 GB on first load. Every later load opens that cache in seconds.

from datasets import load_dataset

# everything
ds = load_dataset("ragarwal/arxiv-markdown", split="train")

# one division: astro-ph, cond-mat, cs, eess, math, physics, quant-ph, q-bio, stat, ...
cs = load_dataset("ragarwal/arxiv-markdown", "cs", split="train")

# one division and year, downloading only those files
cs24 = load_dataset("ragarwal/arxiv-markdown", data_files="cs/2024/*.parquet", split="train")

# metadata only, without the full text
meta = load_dataset("ragarwal/arxiv-markdown", "cs", split="train",
                    columns=["arxiv_id", "title", "abstract", "categories", "published"])

Streaming, without downloading

ds = load_dataset("ragarwal/arxiv-markdown", split="train", streaming=True)

for paper in ds.take(3):
    print(paper["arxiv_id"], paper["title"])

nlp = ds.filter(lambda r: "cs.CL" in r["categories"].split("; "))

Filtering

Batched filters that read only the columns they need are much faster than per-row filters over the full rows.

# papers with cs.CL among their categories
nlp = cs.filter(lambda c: ["cs.CL" in x.split("; ") for x in c],
                input_columns="categories", batched=True, num_proc=8)

# long papers from 2020 on
long = ds.filter(lambda w, y: [a >= 5000 and b >= 2020 for a, b in zip(w, y)],
                 input_columns=["word_count", "year"], batched=True, num_proc=8)

One paper by id

import pyarrow.compute as pc

i = pc.index(ds.data.column("arxiv_id"), "2401.00616").as_py()   # -1 if absent
print(ds[i]["text"][:2000])

Iterating in batches, for example to embed or index

for batch in ds.select_columns(["arxiv_id", "text"]).iter(batch_size=1000):
    ids, texts = batch["arxiv_id"], batch["text"]
    ...

Without datasets

Files are laid out as <division>/<year>/part-NNNN.parquet, so a slice can also be read directly with pandas or polars:

import pandas as pd
df = pd.read_parquet("hf://datasets/ragarwal/arxiv-markdown/cs/2024")

Columns

column	content
arxiv_id	arXiv identifier, e.g. 2401.00637 or hep-lat/9107001
url	abstract page on arxiv.org
title	title
abstract	abstract
authors	author list as given on arXiv
published	timestamp of the first version, UTC
categories	all arXiv categories, primary first, separated by ;
division	archive of the primary category, e.g. cs
year	year of the first version
word_count	whitespace-separated tokens of text
text	the paper as Markdown

Divisions

The division is the archive of a paper's primary category. Papers whose primary category lies outside the 58 but that are cross-listed into one of them keep their own division, which is why hep-th, gr-qc and similar appear.

division	papers
cs	751,463
cond-mat	268,171
physics	167,600
math	160,565
quant-ph	135,161
eess	81,309
stat	61,716
astro-ph	38,998
q-bio	23,221
math-ph	8,866
hep-th	7,615
nlin	4,991
q-fin	4,988
gr-qc	4,595
hep-ph	3,224
econ	3,022
hep-ex	2,046
nucl-th	1,256
hep-lat	1,116
nucl-ex	613

License

Each paper keeps the license its authors chose on arXiv. Most papers use arXiv's non-exclusive distribution license, which grants distribution rights to arXiv only. Check a paper's license on its arXiv page before redistributing its text.