Qwen3-VL-2B OCR Model

This model is a fine-tuned version of Qwen3-VL-2B, optimized for Hindi OCR. It is trained on the iamkushagratomar/devanagari-ocr-synthetic-80k dataset to recognize and extract Hindi text from images.

🔧 How to Use

1. Install dependencies:

   pip install -q -U unsloth unsloth_zoo
   

2. Imports:

    import unsloth
    from unsloth import FastVisionModel
    import torch
    from PIL import Image
   

3. Load the model:

    TEST_IMAGE_PATH = "/content/image.jpg"
    INSTRUCTION = "Read all the text in this image and transcribe it exactly."
    
    inference_model, inference_processor = FastVisionModel.from_pretrained(
        "iamkushagratomar/Qwen3VL-2B-OCR",
        load_in_4bit=True,
    )
    FastVisionModel.for_inference(inference_model)
   

4. Use the model::

   image = Image.open(TEST_IMAGE_PATH).convert("RGB")
   text_prompt = inference_processor.apply_chat_template(
       [
           {"role": "user", "content": [
               {"type": "image", "image": image},
               {"type": "text", "text": INSTRUCTION},
           ]}
       ], 
       tokenize=False, 
       add_generation_prompt=True
   )
   inputs = inference_processor(
       text=[text_prompt], 
       images=[image], 
       padding=True, 
       return_tensors="pt").to(inference_model.device)
   
   with torch.no_grad():
       output_ids = inference_model.generate(**inputs, max_new_tokens=512)
   
   # Example usage
   trimmed = output_ids[:, inputs["input_ids"].shape[1]:]
   result = inference_processor.batch_decode(
       trimmed, 
       skip_special_tokens=True, 
       clean_up_tokenization_spaces=False)[0]
   print("OCR OUTPUT:", result)
   

Uploaded finetuned model

• Developed by: iamkushagratomar
• License: apache-2.0
• Finetuned from model : unsloth/Qwen3-VL-2B-Instruct-unsloth-bnb-4bit

This qwen3_vl model was trained 2x faster with Unsloth