mdnaseif/hafith-synthetic-1m

HAFITH Synthetic Dataset (1M Samples)

Synthetic dataset of 1 million manuscript-style Arabic text line images for training the HAFITH OCR model.

Dataset Summary

• Total Samples: 1,000,000 (900K train / 50K val / 50K test)
• Text Source: ArabicText-Large (244M words)
• Fonts: 350 Arabic fonts (Naskh, Ruq'ah, Thuluth, Maghrebi)
• Image Size: 800×48 to 2500×128 pixels
• Backgrounds: 50 aged parchment variants

Generation Pipeline

1. Text Rendering: Sequential extraction from ArabicText-Large with proper RTL shaping
2. Degradation: Stochastic augmentations simulating manuscript artifacts (paper texture, ink degradation, aging effects)

Key Augmentations

Type	Augmentations	Probability
Geometric	Paper texture, baseline warp	0.50-0.70
Noise	Gaussian noise, blur	0.25-0.30
Ink	Erosion, feathering, bleed	0.15-0.25
Aging	Foxing, stains, show-through	0.10-0.20

Usage

from datasets import load_dataset

# Load dataset
dataset = load_dataset("mdnaseif/hafith-synthetic-1m")

# Or stream for large datasets
dataset = load_dataset("mdnaseif/hafith-synthetic-1m", streaming=True)

# Access sample
sample = dataset['train'][0]
print(sample['text'])
sample['image'].show()

Why Synthetic Data?

Our ablations show that Aranizer tokenization requires synthetic pretraining:

Configuration	CER
NaFlex + Aranizer (no synthetic)	8.47% ❌
NaFlex + Aranizer + Synthetic	5.10% ✅

Without synthetic data, switching to Arabic tokenization degrades performance.

Citation

@article{naseif2026hafith,
  title={HAFITH: Aspect-Ratio Preserving Vision-Language Model for Historical Arabic Manuscript Recognition},
  author={Naseif, Mohammed and Mesabah, Islam and Hajjaj, Dalia and Hassan, Abdulrahman and Elhayek, Ahmed and Koubaa, Anis},
  year={2026}
}

Links

• 🤗 Model: mdnaseif/hafith
• 📊 Real Benchmark: mdnaseif/hafith-combined-benchmark
• 💻 Code: GitHub

License

MIT License - See LICENSE file for details.