SafeVid-350K
收藏arXiv2025-09-30 收录
下载链接：
https://huggingface.co/datasets/yxwang/SafeVid-350K
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是一个包含350,000对视频特定安全偏好记录的新颖数据集，旨在通过强大的文本安全对齐能力，提升视频大型多模态模型（VLMMs）的安全性对齐。该数据集被用于直接偏好优化（DPO），以实现VLMMs的安全对齐，并使用SafeVidBench基准进行评估，以完成视频模型的安全对齐任务。
This novel dataset contains 350,000 pairs of video-specific safety preference records, and is designed to enhance the safety alignment of Video Large Multimodal Models (VLMMs) through robust text-based safety alignment capabilities. It is employed for Direct Preference Optimization (DPO) to achieve the safety alignment of VLMMs, and is evaluated via the SafeVidBench benchmark to complete safety alignment tasks for video models.
