Arabic Synthetic Book Pages for OCR and VLM Training

300,000 synthetic Arabic document pages mimic real-world scanning imperfections for training Vision Language Models and OCR engines. Generated by 'loay', the dataset provides ground truth text in UTF-8 with XML-style annotations for layout analysis and font recognition.

Use Cases

• Train OCR engines on 300,000 synthetic Arabic page images to improve text extraction from degraded scans.
• Fine-tune Vision Language Models using the UTF-8 ground truth text and XML-style layout annotations for document understanding tasks.
• Benchmark font recognition algorithms against the dataset's variety of synthetic Arabic typography and structural layouts.
• Develop robustness models for document degradation by training on the synthetic scanning imperfections present in the ~800×1200 px JPEG images.

Strengths

• Large scale of ~300,000 synthetic Arabic document pages.
• Includes ground truth text in UTF-8 with XML-style annotations for precise training.
• Images are generated to mimic real-world scanning imperfections, adding practical value for model robustness.

Limitations

• Synthetic nature may not fully capture the complexity and noise of real-world, non-simulated document scans.
• Limited information on the diversity of fonts, layouts, and degradation types simulated.
• No explicit temporal or geographic coverage specified for the source material being mimicked.

Provenance

Source

Hugging Face dataset by author 'loay'.

Collection Method

Synthetically generated to mimic real-world scanning imperfections.

Freshness

Last updated on 2026-02-12.