open-index/open-arxiv

Open arXiv

Every arXiv paper's metadata in one place: search, filter, and explore 40 years of science

What is it?

Open arXiv is the complete arXiv metadata dataset, covering titles, abstracts, authors, categories, DOIs, version history, and more. It is converted from the Cornell University Kaggle dataset into Parquet format for efficient querying and streaming.

The dataset contains 2.99M papers spanning from 1991 to 2026, packaged into 417 Parquet shards (Zstd compressed, 1.4 GB total). All 14 fields from the original dataset are preserved with no modifications to the data content.

Released under the CC0 1.0 Universal (Public Domain Dedication) license, the same license as the original dataset.

Dataset at a Glance

Papers	2.99M
Time span	1991 to 2026
Has DOI	43.4%
Revised at least once	39.7%
Average versions	1.6
Cross-listed in multiple categories	47.7%
Parquet shards	417
Total size	1.4 GB

The Growth of arXiv

arXiv started in 1991 as a small physics preprint server at Los Alamos National Laboratory. Today it spans dozens of disciplines. Here is every year of submissions, from the first papers to hundreds of thousands per year:

1991  █ 306
1992  █ 3.3K
1993  █ 6.7K
1994  █ 10.1K
1995  █ 13.0K
1996  ██ 15.9K
1997  ██ 19.6K
1998  ███ 24.2K
1999  ███ 27.7K
2000  ████ 30.6K
2001  ████ 33.2K
2002  █████ 36.1K
2003  █████ 39.4K
2004  ██████ 43.7K
2005  ██████ 46.8K
2006  ███████ 50.2K
2007  ███████ 55.6K
2008  ████████ 58.9K
2009  █████████ 64.0K
2010  █████████ 70.1K
2011  ██████████ 76.6K
2012  ███████████ 84.6K
2013  █████████████ 92.6K
2014  █████████████ 97.5K
2015  ██████████████ 105.3K
2016  ███████████████ 113.4K
2017  █████████████████ 123.5K
2018  ███████████████████ 140.6K
2019  █████████████████████ 155.9K
2020  █████████████████████████ 178.3K
2021  █████████████████████████ 181.6K
2022  ██████████████████████████ 185.7K
2023  █████████████████████████████ 208.5K
2024  ██████████████████████████████████ 244.0K
2025  ████████████████████████████████████████ 284.5K
2026  █████████ 66.8K

SQL query to reproduce this chart

SELECT CASE
    WHEN POSITION('/' IN id) > 0 THEN
      CASE WHEN CAST(SUBSTR(SPLIT_PART(id, '/', 2), 1, 2) AS INT) >= 91
        THEN 1900 + CAST(SUBSTR(SPLIT_PART(id, '/', 2), 1, 2) AS INT)
        ELSE 2000 + CAST(SUBSTR(SPLIT_PART(id, '/', 2), 1, 2) AS INT)
      END
    ELSE 2000 + CAST(SUBSTR(id, 1, 2) AS INT)
  END AS submission_year,
  COUNT(*) AS papers
FROM read_parquet('hf://datasets/open-index/open-arxiv/data/*/*.parquet')
GROUP BY submission_year
ORDER BY submission_year;

What People Write About

The top 20 primary categories, showing where open science is thriving:

#	Category	Papers
1	cs.CV	143.9K	████████████████████
2	hep-ph	141.6K	███████████████████░
3	cs.LG	128.5K	█████████████████░░░
4	quant-ph	128.5K	█████████████████░░░
5	hep-th	112.2K	███████████████░░░░░
6	astro-ph	94.2K	█████████████░░░░░░░
7	cs.CL	78.1K	██████████░░░░░░░░░░
8	gr-qc	70.7K	█████████░░░░░░░░░░░
9	cond-mat.mtrl-sci	69.7K	█████████░░░░░░░░░░░
10	cond-mat.mes-hall	68.7K	█████████░░░░░░░░░░░
11	math.AP	56.6K	███████░░░░░░░░░░░░░
12	astro-ph.GA	53.9K	███████░░░░░░░░░░░░░
13	math.CO	52.3K	███████░░░░░░░░░░░░░
14	cond-mat.str-el	52.2K	███████░░░░░░░░░░░░░
15	astro-ph.SR	47.7K	██████░░░░░░░░░░░░░░
16	astro-ph.HE	45.4K	██████░░░░░░░░░░░░░░
17	astro-ph.CO	44.4K	██████░░░░░░░░░░░░░░
18	math.PR	43.4K	██████░░░░░░░░░░░░░░
19	cond-mat.stat-mech	43.3K	██████░░░░░░░░░░░░░░
20	math.AG	39.6K	█████░░░░░░░░░░░░░░░

SQL query to reproduce this table

SELECT SPLIT_PART(categories, ' ', 1) AS primary_cat,
       COUNT(*) AS papers
FROM read_parquet('hf://datasets/open-index/open-arxiv/data/*/*.parquet')
GROUP BY primary_cat
ORDER BY papers DESC
LIMIT 20;

And grouped by top-level discipline:

#	Category	Papers
1	Computer Science	736.8K	████████████████████
2	Mathematics	571.1K	███████████████░░░░░
3	Condensed Matter	344.9K	█████████░░░░░░░░░░░
4	Astrophysics	332.3K	█████████░░░░░░░░░░░
5	Physics (general)	207.5K	█████░░░░░░░░░░░░░░░
6	High Energy Physics (ph)	141.6K	███░░░░░░░░░░░░░░░░░
7	Quantum Physics	128.5K	███░░░░░░░░░░░░░░░░░
8	High Energy Physics (th)	112.2K	███░░░░░░░░░░░░░░░░░
9	General Relativity	70.7K	█░░░░░░░░░░░░░░░░░░░
10	Electrical Engineering	70.5K	█░░░░░░░░░░░░░░░░░░░
11	Statistics	60.7K	█░░░░░░░░░░░░░░░░░░░
12	Nuclear Physics (th)	35.4K	█░░░░░░░░░░░░░░░░░░░
13	Mathematical Physics	33.9K	█░░░░░░░░░░░░░░░░░░░
14	Quantitative Biology	33.7K	█░░░░░░░░░░░░░░░░░░░
15	High Energy Physics (exp)	24.8K	█░░░░░░░░░░░░░░░░░░░

SQL query to reproduce this table

SELECT CASE
    WHEN POSITION('.' IN SPLIT_PART(categories, ' ', 1)) > 0
      THEN SPLIT_PART(SPLIT_PART(categories, ' ', 1), '.', 1)
    ELSE SPLIT_PART(categories, ' ', 1)
  END AS area,
  COUNT(*) AS papers
FROM read_parquet('hf://datasets/open-index/open-arxiv/data/*/*.parquet')
GROUP BY area
ORDER BY papers DESC
LIMIT 15;

47.7% of papers are cross-listed in more than one category.

Field Coverage

Not every paper fills in every metadata field. Here is how complete each optional field is across the full dataset:

Field	Coverage
submitter	99.5%	██████████████░
authors	100.0%	███████████████
comments	72.6%	██████████░░░░░
journal_ref	31.4%	████░░░░░░░░░░░
doi	43.4%	██████░░░░░░░░░
license	84.9%	████████████░░░
report_no	6.4%	█░░░░░░░░░░░░░░

SQL query to reproduce this table

SELECT
  COUNT(*) AS total,
  ROUND(100.0 * SUM(CASE WHEN submitter IS NOT NULL AND submitter != '' THEN 1 ELSE 0 END) / COUNT(*), 1) AS submitter_pct,
  ROUND(100.0 * SUM(CASE WHEN authors IS NOT NULL AND authors != '' THEN 1 ELSE 0 END) / COUNT(*), 1) AS authors_pct,
  ROUND(100.0 * SUM(CASE WHEN comments IS NOT NULL AND comments != '' THEN 1 ELSE 0 END) / COUNT(*), 1) AS comments_pct,
  ROUND(100.0 * SUM(CASE WHEN journal_ref IS NOT NULL AND journal_ref != '' THEN 1 ELSE 0 END) / COUNT(*), 1) AS journal_ref_pct,
  ROUND(100.0 * SUM(CASE WHEN doi IS NOT NULL AND doi != '' THEN 1 ELSE 0 END) / COUNT(*), 1) AS doi_pct,
  ROUND(100.0 * SUM(CASE WHEN license IS NOT NULL AND license != '' THEN 1 ELSE 0 END) / COUNT(*), 1) AS license_pct,
  ROUND(100.0 * SUM(CASE WHEN report_no IS NOT NULL AND report_no != '' THEN 1 ELSE 0 END) / COUNT(*), 1) AS report_no_pct
FROM read_parquet('hf://datasets/open-index/open-arxiv/data/*/*.parquet');

The required fields (id, title, categories, abstract, versions, update_date, authors_parsed) are present for every paper.

Versions and Revisions

Most papers start at v1 and stay there, but a good number get updates. 39.7% of all papers have been revised at least once. On average, papers have 1.6 versions, and the most-revised paper reached v187.

SQL query to reproduce these stats

SELECT
  ROUND(AVG(json_array_length(versions)), 2) AS avg_versions,
  MAX(json_array_length(versions)) AS max_versions,
  ROUND(100.0 * SUM(CASE WHEN json_array_length(versions) > 1 THEN 1 ELSE 0 END) / COUNT(*), 1) AS revised_pct
FROM read_parquet('hf://datasets/open-index/open-arxiv/data/*/*.parquet');

Licenses

arXiv papers carry their own license terms. Here is the breakdown:

License	Papers	Share
arXiv non-exclusive	1.83M	61.3%
CC BY	502.2K	16.8%
Not specified	452.8K	15.1%
CC BY-NC-ND	78.9K	2.6%
CC BY-NC-SA	59.3K	2.0%
CC BY-SA	27.2K	0.9%
CC0 (Public Domain)	20.0K	0.7%
CC BY	7.9K	0.3%
CC BY-NC-SA	5.9K	0.2%
creativecommons.org/licenses/publicdo...	2.5K	0.1%

SQL query to reproduce this table

SELECT COALESCE(license, '') AS license,
       COUNT(*) AS papers
FROM read_parquet('hf://datasets/open-index/open-arxiv/data/*/*.parquet')
GROUP BY license
ORDER BY papers DESC;

What is being released?

Every arXiv paper's metadata is stored as one row in the Parquet files, organized by submission month:

data/
  1991/
    1991-01.parquet
    1991-02.parquet
    ...
  2007/
    2007-04.parquet
    ...
  2026/
    2026-01.parquet
    2026-02.parquet
    2026-03.parquet

Each shard contains all papers submitted in that month. This layout makes incremental updates efficient: when arXiv is updated weekly, only the current month's shard changes while all historical months stay the same.

Each row includes the paper ID, title, abstract, authors (both raw and parsed), categories, DOI, journal reference, version history, and more. The versions and authors_parsed fields are stored as JSON strings for maximum compatibility.

How to download and use Open arXiv

Using datasets

from datasets import load_dataset

# Stream the entire dataset
ds = load_dataset("open-index/open-arxiv", split="train", streaming=True)
for paper in ds:
    print(paper["id"], paper["title"])

# Load into memory
ds = load_dataset("open-index/open-arxiv", split="train")
print(f"{len(ds):,} papers loaded")

Using huggingface_hub

from huggingface_hub import snapshot_download

folder = snapshot_download(
    "open-index/open-arxiv",
    repo_type="dataset",
    local_dir="./open-arxiv/",
)

For faster downloads, install pip install huggingface_hub[hf_transfer] and set HF_HUB_ENABLE_HF_TRANSFER=1.

Using DuckDB

-- Count papers per top-level category
SELECT split_part(categories, '.', 1) AS area,
       COUNT(*) AS papers
FROM read_parquet('hf://datasets/open-index/open-arxiv/data/*/*.parquet')
GROUP BY area
ORDER BY papers DESC
LIMIT 20;

Working with JSON fields

import json
from datasets import load_dataset

ds = load_dataset("open-index/open-arxiv", split="train", streaming=True)
for paper in ds:
    # Parse structured authors
    authors = json.loads(paper["authors_parsed"])
    for last, first, suffix in authors:
        print(f"  {first} {last}")

    # Parse version history
    versions = json.loads(paper["versions"])
    print(f"  First submitted: {versions[0]['created']}")
    print(f"  Latest version:  v{len(versions)}")

Dataset card for Open arXiv

Dataset Structure

Data Instance

{
  "id": "0704.0001",
  "submitter": "Pavel Nadolsky",
  "authors": "C. Balazs, E. L. Berger, P. M. Nadolsky, C.-P. Yuan",
  "title": "Calculation of prompt diphoton production cross sections at Tevatron and LHC energies",
  "comments": "37 pages, 15 figures; published version",
  "journal_ref": "Phys.Rev.D76:013009,2007",
  "doi": "10.1103/PhysRevD.76.013009",
  "report_no": "ANL-HEP-PR-07-12",
  "categories": "hep-ph",
  "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
  "abstract": "  A fully differential calculation in perturbative quantum chromodynamics is presented for the production of massive photon pairs...",
  "versions": "[{\"version\": \"v1\", \"created\": \"Mon, 2 Apr 2007 19:18:42 GMT\"}, {\"version\": \"v2\", \"created\": \"Tue, 24 Jul 2007 20:10:27 GMT\"}]",
  "update_date": "2008-11-26",
  "authors_parsed": "[[\"Balazs\", \"C.\", \"\"], [\"Berger\", \"E. L.\", \"\"], [\"Nadolsky\", \"P. M.\", \"\"], [\"Yuan\", \"C. -P.\", \"\"]]"
}

Data Fields

Column	Type	Description
id	string	arXiv paper ID (e.g. "0704.0001" or "2301.12345")
submitter	string (nullable)	Person who submitted the paper
authors	string (nullable)	Raw author string exactly as submitted
title	string	Paper title
comments	string (nullable)	Free-form comments (e.g. "37 pages, 15 figures; published version")
journal_ref	string (nullable)	Journal citation (e.g. "Phys.Rev.D76:013009,2007")
doi	string (nullable)	Digital Object Identifier
report_no	string (nullable)	Institutional report number
categories	string	Space-separated arXiv categories (e.g. "hep-ph" or "math.CO cs.CG")
license	string (nullable)	License URL for the paper content
abstract	string	Paper abstract
versions	string (JSON)	Version history array: [{"version": "v1", "created": "..."}]
update_date	string	Last metadata update date (YYYY-MM-DD)
authors_parsed	string (JSON)	Structured authors: [["LastName", "FirstName", "Suffix"]]

Data Splits

The dataset has a single train split containing all 2.99M papers.

Dataset Creation

Curation Rationale

The original arXiv metadata on Kaggle is distributed as a single large JSONL file (about 4-5 GB), which is inconvenient for streaming, SQL queries, and integration with modern ML tooling. Open arXiv converts this to columnar Parquet format with Zstd compression, enabling:

• Streaming via HuggingFace datasets without downloading the full file
• SQL queries via DuckDB directly from HuggingFace URLs
• Column pruning to load only the fields you need (e.g. just titles and abstracts)
• Efficient filtering by category, date, or any other field

Source Data

The source is the Cornell University arXiv dataset on Kaggle, which provides metadata for all papers on arXiv.org. The Kaggle dataset is updated weekly by Cornell University.

Data Processing Steps

1. Download the dataset ZIP from Kaggle (about 1.7 GB compressed)
2. Extract the JSONL file (arxiv-metadata-oai-snapshot.json, about 4-5 GB)
3. Convert to Parquet shards using a streaming Go pipeline, where each line is parsed with gjson for zero-allocation field extraction, then written to Zstd-compressed Parquet via parquet-go
4. Analyze the parquet shards with DuckDB to compute dataset statistics
5. Upload to HuggingFace via the huggingface_hub xet-aware uploader

Field Name Changes

The original Kaggle dataset uses hyphenated field names (journal-ref, report-no). These are converted to underscores (journal_ref, report_no) for compatibility with Parquet column naming conventions and Python attribute access.

Complex Fields

The versions and authors_parsed fields contain nested structures (arrays of objects/arrays) that cannot be directly represented in flat Parquet columns. They are stored as JSON strings. Use json.loads() in Python to parse them.

Considerations for Using the Data

Social Impact

By converting arXiv metadata to an accessible columnar format, we aim to lower the barrier for scientific text mining, citation analysis, and research trend studies. The dataset enables researchers to explore the full history of arXiv without needing Kaggle credentials or parsing large JSON files.

Known Limitations

• This dataset contains metadata only. Paper PDFs and source files are not included. For full-text access, see arXiv on Google Cloud Storage.
• The abstract field may contain leading whitespace and LaTeX notation from the original submissions.
• Author disambiguation is not performed. Use the authors_parsed field for structured name access.

Additional Information

Licensing

The dataset is released under CC0 1.0 Universal (Public Domain Dedication), the same license as the original Kaggle dataset. Individual papers on arXiv have their own licenses, see the license field in each row.

Contact

Please open a discussion on the Community tab for questions, feedback, or issues.

Last Updated

2026-03-24