PEBench
收藏资源简介:
PEBench是一个包含虚构个人实体和相应一般事件场景的合成数据集,由上海人工智能实验室等机构创建,旨在全面评估机器遗忘在多模态大型语言模型中的性能。该数据集由200个虚构个体和40种不同风格的事件场景组成,共计8000张图像。数据集通过合成数据构建,保证了个体和场景之间的一致性和耦合性,可用于评估机器遗忘方法在保护个人隐私和内容安全方面的有效性。
PEBench is a synthetic dataset composed of fictional individual entities and corresponding general event scenarios, created by institutions including the Shanghai AI Laboratory and other relevant organizations. It aims to comprehensively evaluate the performance of machine unlearning in multimodal large language models. This dataset includes 200 fictional individuals and 40 event scenarios with distinct styles, totaling 8000 images. Constructed using synthetic data, it ensures the consistency and coupling between individuals and scenarios, and can be used to assess the effectiveness of machine unlearning methods in protecting personal privacy and content security.




