metadata
base_model:
- Qwen/Qwen2.5-VL-3B-Instruct
datasets:
- yushaohan/ProGuard-data
language:
- en
tags:
- vlm
- safety
- guard
library_name: transformers
pipeline_tag: image-text-to-text
ProGuard-3B
ProGuard is a proactive multimodal safeguard model. It is designed to identify and reason about unknown risks across both text and visual modalities, moving beyond rigid predefined classification systems.
- Arxiv Paper: ProGuard: Towards Proactive Multimodal Safeguard
- Project Page: ProGuard Homepage
- GitHub Repository: ProGuard Implementation, DeepSafe Implementation
This model is the official open-source implementation of ProGuard. For deployment instructions, please refer to this link.
Citation
If you find this model helpful, please cite our research:
@article{yu2025proguard,
title={ProGuard: Towards Proactive Multimodal Safeguard},
author={Yu, Shaohan and Li, Lijun and Si, Chenyang and Sheng, Lu and Shao, Jing},
journal={arXiv preprint arXiv:2512.23573},
year={2025},
url={https://yushaohan.github.io/ProGuard/}
}
@article{zhang2026deepsight,
title={DeepSight: An All-in-One LM Safety Toolkit},
author={Zhang, Bo and Guo, Jiaxuan and Li, Lijun and Liu, Dongrui and Chen, Sujin and Chen, Guanxu and Zheng, Zhijie and Lin, Qihao and Yan, Lewen and Qian, Chen and others},
journal={arXiv preprint arXiv:2602.12092},
year={2026}
}