| 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) |