| # Alef-OCR-Image2Html | |
| An Arabic OCR model that transforms document images,including historical texts, scanned pages, and handwritten materials—into structured and semantic HTML. | |
| ### Key Features | |
| - **Semantic HTML Output:** Generates structured HTML with semantic tags (section, header, main, footer, table, etc.) | |
| - **Multi-format Support:** Handles various document types including historical manuscripts, newspaper articles, scientific papers, invoices, and more | |
| - **Arabic-Optimized:** Fine-tuned specifically for Arabic text recognition and structure extraction | |
| - **Zero-cost Training:** Developed using Kaggle's free tier computational resources | |
| ## Model Architecture | |
| - **Base Model:** Qwen2.5-VL-Instruct | |
| - **Fine-tuning Method:** QLoRA with 4-bit quantization | |
| - **LoRA Configuration:** Rank 16 applied to all modules | |
| - **Optimization:** Unsloth for memory efficiency and training speed | |
| ## Training Data | |
| The model was trained on a custom dataset of **28K image-HTML pairs** consisting of: | |
| - **46% Web-scraped content** (~13K samples): Arabic Wikipedia articles with cleaned semantic HTML | |
| - **54% Synthetic data** (~15K samples): Generated documents mimicking ~13 real-world formats with diverse layouts and styles | |
| For more details, see the [arabic-image2html dataset](https://huggingface.co/datasets/OussamaBenSlama/arabic-image2html). | |
| ## Training Procedure | |
| Training was performed in two stages: | |
| **Stage 1:** | |
| - Data: 40% of training dataset | |
| - Learning rate: 5e-5 | |
| - LR scheduler: Linear | |
| **Stage 2:** | |
| - Data: 30% of training dataset (different split) | |
| - Learning rate: 1e-5 | |
| - LR scheduler: Cosine | |
| ## Performance | |
| Evaluated by the NAMAA community on an anonymous benchmark dataset: | |
| | Model | WER | CER | BLEU | | |
| |-------|-----|-----|------| | |
| | Alef-OCR-Image2Html | 0.92 | **0.72** | **0.19** | | |
| | Qari-OCR-v0.3 (baseline) | **0.84** | 0.73 | 0.17 | | |
| **Key Results:** | |
| - Better Character Error Rate (CER): 0.72 vs 0.73 | |
| - Better BLEU Score: 0.19 vs 0.17 | |
| - Higher Word Error Rate (WER) due to limited diacritics handling in training data | |
| ## Related Resources | |
| - **Dataset:** [arabic-image2html](https://huggingface.co/datasets/OussamaBenSlama/arabic-image2html) | |
| - **Training and Inference Notebooks:** [Available in the repository](https://github.com/OussamaBenSlama/Alef-OCR-Image2Html) | |
| ## Citation | |
| ```bibtex | |
| @misc{alef_ocr_image2html_2025, | |
| title={Alef-OCR-Image2Html: Arabic OCR to Semantic HTML}, | |
| author={Oussama Ben Slama}, | |
| year={2025}, | |
| howpublished={Hugging Face Models}, | |
| url={https://huggingface.co/OussamaBenSlama/Alef-OCR-Image2Html} | |
| } | |
| ``` | |
| ## Acknowledgments | |
| This work builds upon: | |
| - The NAMAA community's state-of-the-art Qari-OCR model | |
| ## License | |
| Apache2.0 | |