ahmedheakl/docatlas_instruct

DocAtlas: Multilingual Document Understanding Across 80+ Languages

DocAtlas is a large-scale, high-fidelity multilingual OCR dataset covering 82 languages and 10 writing systems, built through model-free differential rendering. It provides precise structural annotations in a unified DocTag format encoding layout, text, and component types — without relying on any learned models for core annotation.

This dataset powers the training of DocAtlas-DeepSeek, which achieves state-of-the-art performance (83.37% overall) on multilingual document understanding, outperforming 16 prior models including GPT-4o and Gemini-2.0-Pro.

Key Features

• 82 languages across 10 writing systems (Latin, Cyrillic, Arabic, Devanagari, Chinese, Hangeul, Greek, Abugidas, Syllabaries, Abjads)
• 180,913 training pages with model-free structural annotations
• Dual annotation pipelines: differential rendering of native DOCX documents + synthetic LaTeX-based generation for RTL scripts
• Unified DocTag format jointly encoding component type, geometry, and text content
• 31 structural element types spanning text, tables, headings, figures, equations, charts, and more
• 25+ content domains including Health, Law, Finance, Science, and Education
• DPO-ready: rendering-derived ground truth serves as positive signal for Direct Preference Optimization, achieving stable cross-lingual transfer (+1.9% in-domain, +1.8% out-of-domain) without base-language degradation

Dataset Columns

Column	Type	Description
id	string	Unique identifier for each page sample.
image	image	Rendered page image (the document page as seen by the model).
doctags	string	Structural annotation in the DocTag format — an XML-like schema encoding component types (e.g., <text>, <table>, <section_header>), bounding-box geometry, and text content for each element on the page.
markdown	string	Full page content converted to Markdown, preserving headings, tables, lists, and inline formatting. Suitable for end-to-end page-to-Markdown training and evaluation.
language	string	ISO 639-1 language code (e.g., en, ar, zh, hi) identifying the primary language of the document page.

Usage

from datasets import load_dataset

dataset = load_dataset("ahmedheakl/docatlas_instruct", split="train")

# Access a sample
sample = dataset[0]
print(sample["language"])   # e.g., "en"
print(sample["markdown"])   # Markdown content of the page
print(sample["doctags"])    # DocTag structural annotation
sample["image"].show()      # Display the page image

Streaming (recommended for large-scale loading)

from datasets import load_dataset

dataset = load_dataset("ahmedheakl/docatlas_instruct", split="train", streaming=True)

for sample in dataset:
    print(sample["id"], sample["language"])
    break

Script & Language Coverage

The dataset spans 10 writing systems with broad typological diversity:

Writing System	Example Languages
Latin	English, French, Spanish, Polish, Vietnamese, ...
Cyrillic	Russian, Ukrainian, Serbian, Bulgarian, ...
Arabic	Arabic, Persian, Urdu, Kurdish, ...
Devanagari	Hindi, Marathi, Nepali, ...
Chinese	Chinese (Simplified & Traditional)
Hangeul	Korean
Greek	Greek
Abugidas	Thai, Khmer, Lao, Burmese, Tamil, Telugu, ...
Syllabaries	Japanese, Amharic, ...
Abjads	Hebrew

Code Release

🚧 Data generation and training code will be released soon. Stay tuned!

Citation

If you use DocAtlas in your research, please cite our paper:

@article{heakl2026docatlas,
  title={DocAtlas: Multilingual Document Understanding Across 80+ Languages},
  author={Heakl, Ahmed and Mohamed, Youssef and Sohail, Abdullah and Elbadry, Rania and Nassar, Ahmed and Staar, Peter WJ and Khan, Fahad Shahbaz and Razzak, Imran and Khan, Salman},
  journal={arXiv preprint arXiv:2605.12623},
  year={2026}
}

License

This dataset is released under the CC-BY 4.0 license. All source documents were collected from publicly accessible web content and academic repositories under permissive licenses, with automated PII detection applied to remove sensitive content.