NAMAA-Space/QariOCR-v0.3-markdown-mixed-dataset

QARI Markdown Mixed Dataset

📋 Dataset Summary

The QARI v0.3 Markdown Mixed Dataset is a specialized synthetic dataset designed for training Arabic OCR models with a focus on complex document layouts and HTML structure understanding. This dataset is part of the QARI-OCR project, which achieves state-of-the-art performance in Arabic text recognition.

This dataset contains 37,000 synthetically generated Arabic document images (29.6k train, 3.7k validation, 3.7k test) with corresponding ground truth text in HTML/Markdown format, featuring:

• 🔤 Full diacritical marks (tashkeel) support
• 📝 Mixed font sizes within documents (headers, body text, annotations)
• 🎨 12 distinct Arabic fonts ranging from common Naskh to ornate calligraphic styles
• 📄 Realistic document layouts with structural HTML tags
• 🖼️ Multiple text sources including Basma2423 and YoussefAnwar Arabic news

🎯 Intended Use

This dataset is specifically designed for:

• Training OCR models that need to understand document structure
• Fine-tuning vision-language models for Arabic text recognition
• Developing systems that preserve formatting and layout information
• Research in Arabic document analysis and understanding

📊 Dataset Statistics

Metric	Value
Total Images	37,000
Train Set	29,600 (80%)
Validation Set	3,700 (10%)
Test Set	3,700 (10%)
Text Sources	oddadmix/Basma2423-Text-with-Diacritics-Correction + YoussefAnwar/Arabic-news
Font Variety	12 Arabic fonts
Font Size Range	14px - 100px
Diacritics Support	✅ Full tashkeel
HTML Structure	✅ Preserved
Layout Complexity	✅ High (mixed sizes, headers)

🔧 Data Generation Pipeline

Stage	Process	Details
1. Text Collection	Source gathering	Basma2423 (with diacritics) + YoussefAnwar Arabic news
2. HTML Templating	Layout generation	Mixed font sizes, structural elements
3. Rendering	WeasyPrint → PDF → Image	High-quality document rendering
4. Degradation	Synthetic noise	Clean / Moderate / Heavy variants

📈 Model Performance

When used to train QARI v0.3, this dataset enables:

Metric	Score
Character Error Rate (CER)	0.300
Word Error Rate (WER)	0.485
BLEU Score	0.545
Training Time	11 hours
CO₂ Emissions	1.88 kg eq.

Key Advantages:

• 📐 Superior layout understanding compared to plain text models
• 🏷️ HTML tag preservation for structured document conversion
• ⚡ Resource efficient - 5x less training time than larger datasets
• 🎯 Specialized performance for document structure tasks

Citation

@article{wasfy2025qari,
  title={QARI-OCR: High-Fidelity Arabic Text Recognition through Multimodal Large Language Model Adaptation},
  author={Wasfy, Ahmed and Nacar, Omer and Elkhateb, Abdelakreem and Reda, Mahmoud and Elshehy, Omar and Ammar, Adel and Boulila, Wadii},
  journal={arXiv preprint arXiv:2506.02295},
  year={2025}
}