mdnaseif/hafith-combined-benchmark

HAFITH Combined Benchmark Dataset

Combined benchmark of three Arabic manuscript datasets for OCR evaluation: MUHARAF, KHATT, and RASAM.

Dataset Summary

Dataset	Train	Val	Test	Total	Description
MUHARAF	21,129	1,021	1,278	23,428	Archival documents (19th-20th century)
KHATT	15,886	962	1,197	18,045	Contemporary handwriting (1,000 writers)
RASAM	2,739	915	906	4,560	Maghrebi manuscripts (10th-20th century)
Total	37,203	2,898	3,381	43,482	Combined (after filtering)

Data Fields

• image: Text line image (PIL Image)
• text: Ground truth Arabic text (Unicode normalized)
• source: Dataset origin ('MUHARAF', 'KHATT', or 'RASAM')
• filename: Original filename

Quality Filtering

Dataset has been filtered to remove (2.49% of samples):

• Empty/missing labels
• Mixed Arabic-Latin scripts (>20% Latin)
• Invalid characters
• Non-text images (stamps, seals)
• Multi-line segmentation errors

Usage

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("mdnaseif/hafith-combined-benchmark")

# Load specific split
train = dataset['train']

# Filter by source
muharaf_samples = train.filter(lambda x: x['source'] == 'MUHARAF')

Benchmark Results (HAFITH)

Dataset	CER	WER
MUHARAF	8.35%	24.76%
KHATT	11.21%	37.36%
RASAM	4.95%	18.94%
Combined	5.10%	18.05%

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}
}

Original Datasets

MUHARAF: Al-Zaidy et al. (2024). [NeurIPS 2024]
KHATT: Mahmoud et al. (2014). [Pattern Recognition]
RASAM: Vidal-Gorène et al. (2021). [ICDAR Workshops]

Links

• 🤗 Model: mdnaseif/hafith
• 🔢 Synthetic Data: mdnaseif/hafith-synthetic-1m
• 💻 Code: GitHub

License

Please refer to original dataset licenses. Combined dataset for research purposes only.