---
viewer: false
dataset: unarxive_2024
title: UnarXive 2024
license: mit
language:
  - en
tags:
  - scholarly-nlp
  - information-retrieval
  - citation-network
  - scientific-papers
  - rag
  - qa
size_categories:
  - 1M<n<10M
pretty_name: UnarXive 2024
source_datasets:
  - arXiv.org
dataset_type: text
multilingual: true
---
## Example Preview

Here is a small excerpt of the dataset format:  
[📄 preview.jsonl](./sample/preview.jsonl)


# Dataset Card for UnarXive 2024

**UnarXive 2024** is a large-scale, structured dataset of **2.3 million full-text arXiv papers (1991–2024)**, processed for use in **NLP and information retrieval tasks**. Each paper is provided in a structured JSONL format.


## Features

- **2.28 million structured papers** across physics, CS, mathematics, and other fields
- **Logical section structure** (Introduction, Methods, etc.) for discourse-aware modeling
- **Metadata-rich**: title, authors, abstract, arXiv ID, categories, license, language, cited_by_count
- **In-text citation markers** and linked reference entries
- **30M+ citation links** with **51.2% resolved** to OpenAlex IDs
- **Preserved mathematical expressions**, figures, and tables with positional spans
- **Multilingual content**, mostly in English, but includes French, German, Catalan, and more (25+ languages)


## Usage

The dataset supports tasks such as **citation recommendation**, **scientific QA**, **retrieval-augmented generation (RAG)**, **Citation function and polarity classification** and **paper embedding training**.

## Pipeline 

GitHub: [ines-besrour/UnarXive-2024](https://github.com/ines-besrour/UnarXive-2024)  


---

Dataset licensed under the MIT License.