---
library_name: transformers
tags:
- OCR
base_model:
- Qwen/Qwen2-VL-2B-Instruct
---

# Jawi-OCR-Qwen-v1

## Model Description

This model is a fine-tuned version of [Qwen/Qwen2-VL-2B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct) specialized for Optical Character Recognition (OCR) of historical Malay texts written in Jawi script (Arabic script adapted for Malay language).

### Model Architecture
- **Base Model**: [Qwen2-VL-2B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct)
- **Model Type**: Vision-Language Model
- **Parameters**: 2 billion
- **Language(s)**: Malay (Jawi script)

## Intended Use

### Primary Intended Uses
- OCR for historical Malay manuscripts written in Jawi script
- Digital preservation of Malay cultural heritage
- Enabling computational analysis of historical Malay texts

### Out-of-Scope Uses
- General Arabic text recognition
- Modern Malay text processing
- Real-time OCR applications

## Training Data

### Dataset Description
This was trained and evaluated using 

### Training Procedure
- Hardware used: 1 x H100
- Training time: 6 hours

## Performance and Limitations

### Performance Metrics
- Character Error Rate (CER): 0.86

## How to Use

```python
# Example code for loading and using the model
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
import torch
from qwen_vl_utils import process_vision_info
from PIL import Image

model_name = 'culturalheritagenus/qwen-for-jawi-v1'

model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,  # Use the appropriate torch dtype if needed
    device_map='auto'            # Optional: automatically allocate model layers across devices
)

# Load the processor from Hugging Face Hub
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")

# Add example usage code
image_path = 'path/to/image'
image = Image.open(image_path).convert('RGB')

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": image,
            },
            {"type": "text", "text": "Convert this image to text"},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)

print(output_text)
```

## Citation

```bibtex
@misc{qwen-for-jawi-v1,
  title     = {Qwen for Jawi v1: a model for Jawi OCR},
  author    = {[Miguel Escobar Varela]}, 
  year      = {2024},
  publisher = {HuggingFace},
  url       = {[https://huggingface.co/mevsg/qwen-for-Jawi-v1]},
  note      = {Model created at National University of Singapore }
}
```

## Acknowledgements
Special thanks to [William Mattingly](https://huggingface.co/wjbmattingly), whose finetuning script served as the base for our finetuning approach: https://github.com/wjbmattingly/qwen2-vl-finetune-huggingface