olmOCR-mix-0225

olmOCR-mix-0225 is a dataset of ~250,000 PDF pages which have been OCRed into plain-text in a natural reading order using gpt-4o-2024-08-06 and a special prompting strategy that preserves any born-digital content from each page.

This dataset can be used to train, fine-tune, or evaluate your own OCR document pipeline.

Quick links:

• 📃 Paper
• 🤗 Model
• 🛠️ Code
• 🎮 Demo

Data Mix

Table 1: Training set composition by source

Source	Unique docs	Total pages
Web crawled PDFs	99,903	249,332
Internet Archive books	5,601	16,803
Total	105,504	266,135

Web crawled PDFs are sampled from a set of over 240 million documents crawled from public websites. Books in the Internet Archive set are in the public domain.

Table 2: Web PDFs breakdown by document type

Document type	Fraction
Academic	60%
Brochure	12%
Legal	11%
Table	6%
Diagram	5%
Slideshow	2%
Other	4%

Distribution is estimating by sampling 707 pages, which are classified using gpt-4o-2024-11-20.

Data Format

Each row in the dataset corresponds to a single page, extracted at random, from a source PDF and transformed into plain text.

No source PDF has had more than 3 random pages extracted from it.

Each extracted page is available as a standalone .pdf file, under the pdf_tarballs/ directory.

Features:

{
    'url': string,         # Original URL of the PDF document
    'page_number': int,    # Page number within the document, 1-indexed
    'id': string,          # ID into /pdfs files folder
    'response': {          # OCRed Page information as JSON blob
        'primary_language': string,
        'is_rotation_valid': bool,
        'rotation_correction': int,
        'is_table': bool,
        'is_diagram': bool,
        'natural_text': str   # The actual text of the PDF is here
    }
}

License

This dataset is licensed under ODC-BY-1.0. It is intended for research and educational use in accordance with AI2's Responsible Use Guidelines. The responses were generated from GPT-4o and GPT-4o is subject to OpenAI's terms of use.