| base_model: | |
| - Qwen/Qwen2.5-VL-7B-Instruct | |
| license: apache-2.0 | |
| library_name: transformers | |
| pipeline_tag: video-text-to-text | |
| # PyVision-Video-7B-SFT | |
| [PyVision-RL: Forging Open Agentic Vision Models via RL](https://arxiv.org/abs/2602.20739) | |
| This is **PyVision-Video-7B-SFT**, post-trained from [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct). | |
| - **Project Page:** [https://agent-x.space/pyvision-rl/](https://agent-x.space/pyvision-rl/) | |
| - **Repository:** [https://github.com/agents-x-project/PyVision-RL](https://github.com/agents-x-project/PyVision-RL) | |
| - **Paper:** [arXiv:2602.20739](https://arxiv.org/abs/2602.20739) | |
| ## Model Description | |
| PyVision-Video is part of the PyVision-RL framework, which aims to stabilize Reinforcement Learning (RL) training for open-weight multimodal models to sustain agentic interaction. | |
| For video reasoning, PyVision-Video employs an **on-demand context construction** strategy. It selectively samples task-relevant frames during the reasoning process, which significantly reduces visual token usage while maintaining strong performance on complex video understanding tasks. This model serves as the Supervised Fine-Tuning (SFT) checkpoint before RL training. | |
| ## Citation | |
| If you find this work useful, please cite the following paper: | |
| ```bibtex | |
| @article{pyvisionrl2026, | |
| title={PyVision-RL: Forging Open Agentic Vision Models via RL}, | |
| author={Zhao, Shitian and Lin, Shaoheng and Li, Ming and Zhang, Haoquan and Peng, Wenshuo and Zhang, Kaipeng and Wei, Chen}, | |
| journal={arXiv:2602.20739}, | |
| year={2026} | |
| } | |
| ``` |