sherif1313/Arabic-English-handwritten-OCR-Qwen3-VL-4B

💜 Github   |   🤗 Hugging Face   |   📚 Cookbooks  
🖥️ Demo  

Knowledge Transfer for Improved Arabic-English Handwritten OCR with Qwen3-VL

Note: This model is intended for research purposes only and is not recommended for production deployment without significant optimization and testing.

Overview

This model enhances Optical Character Recognition (OCR) for sherif1313/Arabic-English-handwritten-OCR-Qwen3-VL-4B texts with high accuracy. The model features successful knowledge transfer from previous versions sherif1313/Arabic-English-handwritten-OCR-v3, providing a strong foundation for further customized training.

🔄 Knowledge Transfer

📊 Transfer Statistics

Metric	Value	Interpretation
Transfer Ratio	95%	Percentage of successfully transferred weights
Time Savings	90%	Compared to training from scratch
Eval Loss	17%	From 0.45 to 0.2982
Performance Impact	-10%	Lower than the old model
Processing Speed	+400%	Slower than the old model

Key Features:

Knowledge transfer from Qwen2.5-VL model to Qwen3-VL
Efficient adaptation using LoRA (Low-Rank Adaptation) technique
Optimized for Arabic and English handwritten text recognition
Scalable for additional OCR tasks

🛠️ Layer Mapping

# Mapping model layers between versions
TARGET_MODULES_MAPPING = {
    "language_model.model.layers.{layer}.self_attn.q_proj": 
        "model.layers.{layer}.self_attn.q_proj",
    "language_model.model.layers.{layer}.self_attn.v_proj": 
        "model.layers.{layer}.self_attn.v_proj",
    "language_model.model.layers.{layer}.mlp.gate_proj": 
        "model.layers.{layer}.mlp.gate_proj",
    "language_model.model.layers.{layer}.mlp.up_proj": 
        "model.layers.{layer}.mlp.up_proj",
    "language_model.model.layers.{layer}.mlp.down_proj": 
        "model.layers.{layer}.mlp.down_proj"
}

Model Limitations###

Application Performance:
    400% slower than previous version
    10% lower accuracy in practical applications

WARNING: This model is for research only

1. ❌ Not designed for production use
2. ⚠️ Performance is lower than previous versions, but it's better for handwriting in English.
3. 🐌 Very slow inference speed
4. 🔬 For academic and research purposes only
5. 📚 Case study in knowledge transfer between generations

Suggested Research Areas

Knowledge Transfer Optimization:
    Study impact of structural mismatch
    Develop automatic transfer tools
    Analyze information loss during transfer
Performance Enhancement:
    Reduce inference time
    Increase accuracy on Arabic text
    Improve handling of diverse handwriting styles
Task Expansion:
    Number and symbol recognition
    Multi-page document processing
    Support for additional languages

Support and Contribution

Reporting Issues

Open an Issue in the model repository
Include:
    Error text
    Usage code
    Runtime environment
    Input examples

Contributing to Development

Contributions are welcome for:

Performance optimization
Additional language support
Documentation improvements
Bug fixes

Models and Data

Qwen3-VL-4B-Instruct
 - Base model
Arabic-English-handwritten-OCR-v3
 - Knowledge source
PEFT 
 - Efficient adaptation 
 

🏷️ Citation

If you use this model in your research or project, please cite:

@software{Arabic-English-handwritten-OCR,
  title={Arabic-English Handwritten OCR v3 (Qwen3-VL)},
  author={sherif1313},
  year={2026},
  url={https://huggingface.co/sherif1313/Arabic-English-handwritten-OCR-Qwen3-VL-4B}
}