Datasets:

nick007x
/

arxiv-papers

license: mit
language:
  - en
size_categories:
  - 1T<n<10T
task_categories:
  - text-to-image
  - visual-question-answering
  - document-question-answering
  - text-generation

Complete ArXiv Papers Dataset (4.68 TB)

📚 Dataset Overview

This repository contains the complete ArXiv scientific papers archive organized by subject categories and publication years. With 4.68 TB of compressed PDFs and metadata, this represents one of the largest collections of scientific literature available for research and AI training.

🗂️ Dataset Structure

Organized by Subject Categories:

astro-ph (00-22): Astrophysics
cond-mat (00-32): Condensed Matter Physics
cs (00-62): Computer Science (most extensive category)
math (00-52): Mathematics
physics (00-16): General Physics
quant-ph (00-12): Quantum Physics
stat (00-05): Statistics
econ, eess, hep, nlin, q-bio, q-fin: Specialized categories
Plus additional specialized domains

File Organization:

Each category split into numbered segments (00, 01, 02...)
Large categories further divided into parts (part-1, part-2, etc.)
All files in ZIP format containing PDFs

📊 Dataset Statistics

Total Size: 4.68 TB (compressed)
Format: ZIP archives containing PDFs + metadata
Coverage: Complete ArXiv historical archive
Organization: By subject category

🎯 Primary Use Cases

Multi-Modal AI Training

Scientific Document Understanding: Train models on full PDF content
Figure-Caption Alignment: Extract and pair scientific figures with their descriptions
Mathematical Reasoning: Process complex mathematical notation and derivations
Cross-modal Retrieval: Link textual concepts with visual scientific content

Research Applications

Bibliometric Analysis: Track research trends across decades
Scientific NLP: Train domain-specific language models
Knowledge Extraction: Parse algorithms, methodologies, and results
Academic Search: Build enhanced scientific search engines

🛠️ Usage Examples

Accessing Specific Categories

# Example: Access Computer Science papers from segment 00
from huggingface_hub import hf_hub_download
import zipfile

file_path = hf_hub_download(
    repo_id="nick007x/xiv-papers",
    filename="cs-00.zip"
)

# Extract and process PDFs
with zipfile.ZipFile(file_path, 'r') as zip_ref:
    zip_ref.extractall("cs_papers/")

Working with Metadata

The train.parquet file contains structured metadata including:

arXiv IDs, titles, authors, submission dates
Abstracts, comments, primary subjects
File paths to corresponding PDFs

⚡ Quick Start

Browse Categories: Start with smaller categories like gr-qc-00.zip (4.07 GB)
Extract Metadata: Use train.parquet for paper discovery
Targeted Download: Download specific subject areas of interest
Stream Processing: Handle large files with streaming extraction

🌟 Value Proposition

This dataset enables:

Complete Scientific Coverage: Every paper from ArXiv's history
Multi-Domain Expertise: Physics, CS, Math, Statistics, and more
Ready for Foundation Models: Perfect for training scientific AI
Structured Organization: Easy access by domain and time period

📜 License & Usage Terms

Important: This dataset is a collection of individual scholarly works from arXiv.org. The licensing structure is as follows:

The Collection (Metadata & Packaging): The script used to create this dataset, the unique metadata (e.g., file structure, dataset description), and its packaging are licensed under the MIT License.
The Individual Papers (Content): Each paper (PDF/TeX source) remains under the copyright and license chosen by its respective author(s). These licenses are typically Creative Commons (e.g., CC BY, CC BY-NC, CC BY-NC-ND).
Your Responsibility: Users of this dataset are solely responsible for checking, understanding, and complying with the specific license terms of any paper they access, download, or use from this collection. You must provide appropriate attribution to the original authors as required by their chosen license.

By using this dataset, you agree to these terms and acknowledge that the dataset creator is not liable for any license violations resulting from your use of the contained papers.

For more information, see:

🙏 Acknowledgments

This dataset builds upon the incredible work of:

ArXiv team and moderators
Paper authors across all scientific domains
Open scientific community enabling knowledge sharing

Note: Due to the massive size (4.68 TB), consider downloading specific categories of interest rather than the entire dataset. The organized structure makes targeted access straightforward.