ArXiv-Papers-150K

150K+ ArXiv papers as raw LaTeX source archives, covering major AI/ML conferences (2016--2026).

Overview

Papers	150,334
Size	~285 GB
Format	.tar.gz per paper (original ArXiv source)
Years	2016 -- 2026
Categories	cs.LG, cs.CV, cs.CL, cs.AI, stat.ML, cs.NE, cs.SD, eess.AS, cs.RO

Category Breakdown

Category	Papers	Description
cs.LG	54,200	Machine Learning (ICML, NeurIPS, ICLR)
cs.CV	35,000	Computer Vision (CVPR, ECCV, ICCV)
cs.CL	25,000	Computation & Language (ACL, EMNLP)
cs.AI	20,000	Artificial Intelligence (AAAI, IJCAI)
stat.ML	7,803	Statistical Machine Learning
cs.NE	2,500	Neural & Evolutionary Computing (IJCNN)
cs.RO	2,500	Robotics
cs.SD	1,500	Sound
eess.AS	1,500	Audio & Speech

Dataset Structure

repo/
  metadata.parquet              # Paper metadata (title, abstract, authors, categories, year, chunk_file)
  cs.LG/
    cs.LG_part_000.tar          # ~10GB each, contains individual paper .tar.gz files
    cs.LG_part_001.tar
    ...
  cs.CV/
    cs.CV_part_000.tar
    ...

Each .tar chunk contains individual paper archives (.tar.gz). Each paper archive contains the LaTeX source as submitted to ArXiv: .tex files, figures, .bib references, style files, etc.

Metadata

metadata.parquet contains the following columns for all 150K papers:

Column	Type	Description
paper_id	string	ArXiv ID (e.g., 2401.12345)
title	string	Paper title
abstract	string	Abstract
authors	string	Comma-separated author names
categories	string	All ArXiv categories
primary_category	string	Primary ArXiv category
published	string	Publication date
year	int	Year extracted from ArXiv ID
chunk_file	string	Which tar chunk contains this paper
doi	string	DOI if available
journal_ref	string	Journal reference if available

Quick Start

Load metadata:

import pandas as pd
df = pd.read_parquet("hf://datasets/Vidushee/ArXiv-Papers-150K/metadata.parquet")
print(df.shape)  # (150334, 12)

Download and extract a category:

from huggingface_hub import hf_hub_download
import tarfile, os

# Download one chunk
path = hf_hub_download(
    repo_id="Vidushee/ArXiv-Papers-150K",
    filename="cs.LG/cs.LG_part_000.tar",
    repo_type="dataset",
)

# Extract paper archives
with tarfile.open(path) as tar:
    tar.extractall("./cs_LG_papers/")

# Each file is a paper's source archive
# Extract a single paper
paper = "./cs_LG_papers/2401.12345.tar.gz"
with tarfile.open(paper) as t:
    t.extractall("./paper_source/")
# Now you have: main.tex, figures/, references.bib, etc.

Use Cases

• LLM pre-training on scientific text
• Scientific document understanding
• LaTeX generation and completion
• Citation network analysis
• Figure extraction and captioning

Source

All papers sourced from arxiv.org via the /e-print/ endpoint. Metadata from anonymousatom/arxiv-metadata. Papers distributed under their original ArXiv licenses.

Citation

@dataset{arxiv_papers_150k,
  title={ArXiv-Papers-150K: LaTeX Source Archives for 150K AI/ML Papers},
  year={2026},
  url={https://huggingface.co/datasets/Vidushee/ArXiv-Papers-150K},
  note={150K ArXiv paper source archives covering cs.LG, cs.CV, cs.CL, cs.AI, stat.ML and related categories (2016-2026)}
}