mdnaseif/hafith-synthetic-1m

---
language:
- ar
task_categories:
- image-to-text
tags:
- ocr
- arabic
- manuscripts
- synthetic-data
size_categories:
- 100K<n<1M
license: mit
---

# 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

```python
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

```bibtex
@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](https://huggingface.co/mdnaseif/hafith)
- 📊 **Real Benchmark**: [mdnaseif/hafith-combined-benchmark](https://huggingface.co/datasets/mdnaseif/hafith-combined-benchmark)
- 💻 **Code**: [GitHub](https://github.com/mdnaseif/hafith)

## License

MIT License - See LICENSE file for details.