SafeEdit
收藏资源简介:
浙江大学与蚂蚁集团等机构合作构建的SafeEdit数据集,旨在探讨知识编辑技术对大型语言模型解毒的可行性。该数据集包括4,050个训练样本、2,700个验证样本和1,350个测试样本,覆盖九种不安全类别,如法律、政治敏感、伦理道德等,并提供强大的攻击提示以诱导模型产生有害内容。SafeEdit的设计考虑了模型对恶意输入的防御成功率、对新型恶意输入的泛化防御能力,以及模型对正常用户请求的响应能力。该数据集为研究LLM的排毒问题提供了新的数据和评估框架,有助于推动该领域的发展,并可以应用于多种方法,如监督式微调、强化学习和知识编辑方法。
The SafeEdit dataset, collaboratively constructed by Zhejiang University, Ant Group and other institutions, is designed to investigate the feasibility of detoxifying large language models (LLMs) using knowledge editing techniques. Comprising 4,050 training samples, 2,700 validation samples, and 1,350 test samples, this dataset covers nine unsafe categories including legal-sensitive, politically-sensitive, ethical and moral issues, etc., and provides robust attack prompts to induce models to generate harmful content. The design of SafeEdit considers three key aspects: the defense success rate of models against malicious inputs, the generalized defense capability against novel malicious inputs, and the model's ability to respond to normal user requests. This dataset offers novel data resources and an evaluation framework for research on LLM detoxification, facilitating the advancement of this research field, and is applicable to multiple methodologies such as supervised fine-tuning, reinforcement learning, and knowledge editing approaches.




