README.md

---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
---

# SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models


This model, VLAA-Thinker-Qwen2VL-7B, is a vision-language model fine-tuned on the VLAA-Thinking dataset. As described in [](https://huggingface.co/papers/2504.11468), it leverages a combination of supervised fine-tuning (SFT) and reinforcement learning (RL) to improve reasoning capabilities in LLMs. The model excels in multimodal reasoning tasks, achieving state-of-the-art performance on the OpenCompass Multimodal Reasoning Leaderboard as of April 7th, 2025.

<p align="center">
  🌐 <a href="https://ucsc-vlaa.github.io/VLAA-Thinking/" target="_blank">Project Page</a>  
  • <img src="./assets/ar.svg" alt="Arxiv Logo" style="height: 1em; vertical-align: middle; margin-right: 0.3em;">
  <a href="./assets/VLAA-Thinker.pdf" target="_blank">Arxiv</a>  
  • 💻  <a href="https://github.com/UCSC-VLAA/VLAA-Thinking" target="_blank">Code</a>  
</p>


Both **VLAA-Thinker-Qwen2.5-3B** and **VLAA-Thinker-Qwen2.5-7B** achieve **SOTA** performance on [OpenCompass Multimodal Reasoning Leaderboard](https://rank.opencompass.org.cn/leaderboard-multimodal-reasoning/?m=REALTIME) as of April 7th, 2025.
<img src="assets/opencompass_4b_box.png" width = "640" alt="pipeline" align=center />

-----

<img src="assets/opencompass_7b_box.png" width = "640" alt="pipeline" align=center />




## Quick Start 🚀
### Inference
Run `python inference.py`. Note that our model is trained with a system prompt. Please ensure that it is included for inference.


### Dataset Download
Run `bash ./utils/download_dataset.sh`. Specify the dataset root with absolute path. The dataset should be ordered as follows:
```
├── VLAA-Thinking-SFT-126K.json
├── VLAA-Thinking-GRPO-25K.json
└── images
    ├── allava_laion
    ├── arxivqa
    ├── chartqa
    ├── clevr_math
    ├── coco
    │   └── train2017
    ├── docvqa
    ├── geoqa170k
    ├── synthesis
    ├── vg
    │   ├── VG_100K
    │   └── VG_100K_2
    └── vizwiz
```
### Training
Code coming soon!

(Rest of the README content can be kept as is)