ATPO-G-Qwen2.5-VL-7B

ATPO-G-Qwen2.5-VL-7B is based on Qwen2.5-VL-7B-Instruct and fine-tuned for multi-label video safety detection using Adaptive Tversky Policy Optimization (ATPO). The ATPO-G variant uses Global Adaptive Tversky Reward (G-ATR), which adapts false-positive and false-negative penalties during reinforcement learning to steer the precision-recall trade-off.

This repository contains merged BF16 model weights, tokenizer files, and processor configurations.

• Paper: Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization
• Code: BruceYg/ATPO
• Project page: ATPO

Citation

@inproceedings{yang2026controllable,
  title={Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization},
  author={Guangyu Yang and Jingbiao Mei and Mingsheng Sun and Jinghong Chen and Yingtong Bu and Pengda Qin and Da Chen and Bill Byrne},
  booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
  year={2026},
  url={https://arxiv.org/abs/2610.02019}
}