README.md

---
library_name: transformers
pipeline_tag: image-text-to-text
license: mit
tags:
  - multimodal
  - vision-language
  - reasoning
  - qwen2
---

# Model Card for Virgo-72B

Virgo is a multi-modal slow-thinking reasoning model based on Qwen2-VL-72B-Instruct. It excels in image-text-to-text tasks, demonstrating strong performance on various multimodal benchmarks.  Virgo leverages a long-form thought process for enhanced reasoning capabilities, effectively integrating visual information into its responses.

## Model Details

### Model Sources

- **Repository:** https://github.com/RUCAIBox/Virgo
- **Paper:** https://arxiv.org/pdf/2501.01904

## Quick Start

This example demonstrates how to use Virgo-72B with the `vllm` library for text generation given an image and text input. Ensure you have `vllm` and `Pillow` installed (`pip install vllm Pillow`) and a suitable image file (`case/2246_image_1.jpg` in this example).

```python
from vllm import LLM, SamplingParams
from PIL import Image

model_name = "RUC-AIBOX/Virgo-72B"
placeholder = "<|image_pad|>"
llm = LLM(
        model=model_name,
        trust_remote_code=True,
        tensor_parallel_size=8,  # Adjust based on your hardware
    )
question = "Please first think deeply about the question, and then put the final answer in \\boxed{}.
In the diagram, $\\angle E A D=90^{\\circ}, \\angle A C D=90^{\\circ}$, and $\\angle A B C=90^{\\circ}$. Also, $E D=13, E A=12$, $D C=4$, and $C B=2$. Determine the length of $A B$."
prompt = ("<|im_start|>system
You are a helpful assistant.<|im_end|>
"
              f"<|im_start|>user
<|vision_start|>{placeholder}<|vision_end|>"
              f"{question}<|im_end|>
"
              "<|im_start|>assistant
")
sampling_params = SamplingParams(
    temperature=0.0,
    top_k=1,
    top_p=1.0,
    repetition_penalty=1.05,
    max_tokens=8192
)
image = Image.open("case/2246_image_1.jpg")
inputs = {
            "prompt": prompt,
            "multi_modal_data": {
                "image": image
            },
        }
outputs = llm.generate(inputs, sampling_params)
print(outputs[0].outputs[0].text)
```

## Citation

```
@article{du2025virgo,
      title={Virgo: A Preliminary Exploration on Reproducing o1-like MLLM}, 
      author={Yifan Du and Zikang Liu and Yifan Li and Wayne Xin Zhao and Yuqi Huo and Bingning Wang and Weipeng Chen and Zheng Liu and Zhongyuan Wang and Ji-Rong Wen},
      journal={arXiv preprint arXiv:2501.01904},
      year={2025}
}
```