license: mit
task_categories:
- question-answering
- text2text-generation
- text-generation
language:
- en
pretty_name: uparxive
size_categories:
- 1M<n<10M
configs:
- config_name: example_data
data_files: example.json
The Uparxive Dataset
The uparxive aims to provide a llm-friendly dataest for the whole arxiv .tex source. A similiar dataset is the unarxive, while the uparxvie use a different tool chain.
The Uparxive dataset is stored in .json
format, which can be seamlessly converted into Markdown .md
format.
See github repo for a simple converter.
Format Rules
The Uparxive dataset adheres to the following rules:
Tables and Figures: Elements enclosed within
\begin{table} \end{table}
and\begin{figure} \end{figure}
tags are extracted and appended at the end of the document for clarity and organization.Citations and References: Citations (
\cite{}
) and references (\ref{}
) are converted to more explicit forms to improve readability. Examples include:- Direct mentions:
(See [Ref. [1,2] of ArXiv.1512.03385])
- Contextual references:
in [Ref. [1,2] of ArXiv.1512.03385]
- Equation/Section/Figures/Tables references:
in [Equation [1] of ArXiv.1512.03385]
, depending on the usage context.
- Direct mentions:
Mathematical Notations:
- In-line Math: Single dollar signs
$
are used for in-line mathematical expressions, e.g.,$\alpha$
. - Block Math: Double dollar signs
$$
denote block mathematical expressions, e.g.,$$\mathbf{y}=\mathcal{F}(\mathbf{x},\{W_{i}\})+\mathbf{x}.$$
- In-line Math: Single dollar signs
See 1512.03385.json and 1512.03385.md as example
Dataset Sources [optional]
To construct the complete Uparxive dataset from scratch, please refer to the GitHub repo.
Note: The "full" version of uparxive was generated from all of arXiv.org including non-permissively licensed papers. Make sure that your use of the data is compliant with the paper's licensing terms. (For information on papers' licenses use arXiv's bulk metadata access).
Note: paper under CC BY-NC-ND license are not included in the dataset. [More Information Needed]
Who are the source data producers?
- arXiv Bulk Data Access: Access and download bulk data directly from arXiv using the AWS S3
request-payer
dataset. Detailed instructions and access points can be found here: arXiv Bulk Data Access. - arXiv API: For more specific data needs or metadata, use the arXiv API. Documentation and usage guidelines are available here: arXiv API.
- Important Note: When crawling arXiv source files, ensure to use
export.arxiv.org
instead of the officialarxiv.org
domain to avoid overloading the main site.
- Important Note: When crawling arXiv source files, ensure to use
[More Information Needed]
Statistic
This dataset part will need 156G dick space for 1.7M papers.
Up to April 2024, there are around 2,450,893 papers in the arxiv source, and the uparxive dataset has covered 1,700,000 papers. Those missing parts are mainly due to the lack of the .tex
source or the failure of the conversion process.