---
license: mit
datasets:
- nielsr/docvqa_1200_examples_donut
language:
- en
library_name: transformers
pipeline_tag: visual-question-answering
---

### IDEFICS2-OCR

Finetuned of Idefics2-8b with fp16 weight update on nielsr/docvqa_1200_examples_donut dataset for document VQA pairs.

## Usage

```Python
from transformers import BitsAndBytesConfig, AutoModelForVision2Seq, AutoProcessor
from transformers.image_utils import load_image

processor = AutoProcessor.from_pretrained("smishr-18/Idefics2-OCR", do_image_splitting=False)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = AutoModelForVision2Seq.from_pretrained(
    "smishr-18/Idefics2-OCR",
    quantization_config=bnb_config,
    device_map=device,
    low_cpu_mem_usage=True
    )

image = load_image("https://images.pokemontcg.io/pl1/1_hires.png")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Explain."},
            {"type": "image"},
            {"type": "text", "text": "What is the reflex energy in the image?"}
        ]
    }
]

text = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=[text.strip()], images=[image4], return_tensors="pt", padding=True)
inputs = {k: v.to(device) for k, v in inputs.items()}

# Generate texts
generated_ids = model.generate(**inputs, max_new_tokens=500)
generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(generated_texts)
# The reflex energy in the image is 70.
```

## Limitations

The model was finetuned on limited T4 GPU and could be fintuned with more adapters on 
devices with ```torch.cuda.get_device_capability()[0] >= 8``` or Ampere GPUs.

- **Developed by:** Shubh Mishra, Aug 2024
- **Model Type:** VLM
- **Language(s) (NLP):** English
- **License:** MIT
- **Finetuned from model:** HuggingFaceM4/idefics2-8b