Makadi86/

Arabic-GLM-OCR-v2-backup

💜 Github   |   🤗 Hugging Face   |   📚 Cookbooks  
🖥️ Demo  

A powerful Arabic OCR model (proficient learner)

This model is an advanced Arabic OCR system designed to combine deep linguistic understanding with high accuracy in visual text extraction.

The model was trained using a unique strategy focused on:

Reducing the model's active capacity during training Maintaining the stability of visual features Promoting genuine language understanding rather than rote memorization

🔹 Model size: Approximately 2 GB 🔹 Performance: Outperforms much larger models in most tasks 🔹 Type: Robust learning model (requires fine-tuning for inference)

✅ Deep understanding of Arabic language context ✅ Intelligent spelling correction ✅ High visual accuracy in text extraction ✅ Noise reduction ✅ Highly stable training behavior ✅ Strong generalization on non-visual data 🧪 Evaluation Results Metric Value Evaluation loss 0.1041 Training-evaluation gap 0% - 2.5% Excellent stability

The model was trained using only half its capacity in order to:

Preserve visual representations Prevent image deterioration Improve overall stability 2. From "Memorizing Shapes" to "Learning Rules"

Instead of:

Memorizing word shapes

The model now learns:

Grammar rules and image-text relationships

The training included:

Reducing excessive inference Limiting the linking of complex ideas Reverting processed information to its original size before output

Forcing the model to accurately copy text instead of paraphrasing it

The model was given internal inference capabilities during:

Reading the page Analyzing the text Generating output

This leads to:

Better understanding of invisible data Stronger real-world performance ⚙️ Inference Settings (Very Important)

⚠️ This is a powerful learner ← Requires precise control during inference

🎯 Use Cases 📄 OCR for Arabic books 📰 Text extraction from images 📚 Manuscript digitization 🧾 Document processing 🔍 Text enhancement after OCR ⚠️ Important Notes The model may attempt autocorrect if not properly constrained. To accurately copy text, use directives such as: Extract the text exactly as it is, without correction or paraphrasing.

📦 Why is the model small?

Despite its small size (approximately 2 GB), its outstanding performance is due to:

Effective training methodology Minimized cognitive noise Focus on patterns Significant Highly Efficient Representation Learning 🏁 Conclusion

This model achieves a rare balance between:

Visual Accuracy 👁️ Language Comprehension 🧠 Training Stability ⚖️

💡 It can be considered a sophisticated model for Arabic OCR, competing with larger systems.

In some cases, the model may attempt to correct the text if it is not properly configured. For exact copying: Use a clear prompt such as: "Extract the text as is, without modification"

❌ Do not use high temperature settings → will cause hallucinations. ✅ Use "Restricted" settings for optimal accuracy. ✅ Best suited for OCR tasks, not creative writing. Send feedback Press tab for actions

Recommended Settings It includes:

with torch.no_grad():

generated_ids = model.generate( **inputs, max_new_tokens=512, # Keep repeating the loop do_sample=True, temperature=0.4, top_p=0.9, repetition_penalty=1.1

git clone https://github.com/zai-org/glm-ocr.git cd glm-ocr uv venv --python 3.12 --seed && source .venv/bin/activate uv pip install -e .