Dataset card for arxiv-markdown

[25/04/2025] Images are now hosted on Cloudflare R2 and referenced in the markdowns as external URLs rather than embedded base64. This significantly reduces storage requirements while maintaining full image access. Images are referenced in the dataset as . Community mirrors are welcome!

This dataset contains open-access papers retrieved from arXiv and converted to markdown format using docling; specifically, we use the following docling pipeline:

pipeline_options = PdfPipelineOptions()
pipeline_options.images_scale = IMAGE_RESOLUTION_SCALE  # 2.0
pipeline_options.generate_page_images = True  # Necessary for the extraction of figures (as far as I understand)
pipeline_options.generate_picture_images = True  # For the extraction of figures
pipeline_options.do_code_enrichment = True  # For obtaining code blocks
pipeline_options.do_formula_enrichment = True  # For converting formulas to LaTeX
converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options)
    }
)

Currently the dataset constains entries extracted from August 2014, August 2019 and August 2024; we are currently dedicating an RTX 3090 full-time for the extraction of this data and will continue to upload new entries as they get processed.

You can find the code we are using for the data generation on GitHub (optimizations and general suggestions are very welcome!).

Processed Papers

The distribution of the data currently is like this:

• 2014-08: 1097 papers (and growing)
• 2019-08: 994 papers
• 2024-08: 1178 papers

More papers will be added for already present months and more!

Dataset Scope

This dataset is being built as the first step to expanding the academic-chains dataset, but other creative uses of this dataset by the community are welcome and encouraged! You can check our license below in this same file.

Limitations and Biases

• Extraction Fidelity: While docling is amazing, it is not perfect, and extraction glitches (especially in tables) may still be present
• Slow Data Generation: Document extraction while doing formula and code enrichment and picture extraction is SLOW on our 3090; we didn't use these options while extracting MDs for the academic-chains dataset because it was not strictly necessary, but we would like to do things right this time, even though it will slow us down (relatively)

Note: We see there is work going on on supporting batched inference for docling, and updates on using VLMs too, so we will try to keep our pipeline up to date and efficient!

Acknowledgements

A big-big-BIG thank you to arXiv and to all the authors of the open-access papers present in our dataset (and to all the others too!). And thank you to whoever supported my original academic-chains dataset too, you are the reason why I started working on this so soon and with a smile on my face!

Licensing Information

This dataset is licensed under the CC-BY-4.0 License.

Citation Information

@misc{marcodsn_2025_arxivmarkdown,
    title = {arxiv-arkdown},
    author = {Marco De Santis},
    month = {April},
    year = {2025},
    url = {https://huggingface.co/datasets/marcodsn/arxiv-markdown}
}