VSCBench

该数据集包含了3600组图像-文本对，这些对子在视觉或文本上相似，但在安全性方面存在差异。这一新颖的设计旨在评估在以图像为中心和以文本为中心的场景下，视觉-语言模型的安全校准能力。该数据集旨在解决视觉-语言模型中存在的安全性不足和过度安全的问题，并有助于在各种模型中评估安全校准的性能。规模上，该数据集包含了3600对图像-文本组合，其任务是对视觉-语言模型的安全校准进行评估。

This dataset contains 3600 image-text pairs that are similar either visually or textually but differ in terms of safety. This novel design aims to evaluate the safety calibration capability of vision-language models (VLMs) in both image-centric and text-centric scenarios. This dataset is developed to address the issues of insufficient safety and over-safety in vision-language models, and facilitate the evaluation of safety calibration performance across various models. With a total of 3600 image-text pairs, its core task is to assess the safety calibration of vision-language models.