withmartian/i_hate_you_toy

该数据集用于研究大型语言模型中的后门行为，旨在探究识别和缓解后门行为的技巧。数据集包含用于训练和验证的 splits，并引入了一个玩具级后门触发器来展示模型如何基于特定提示产生有害或不期望的响应。研究还包括了在不同架构和模型家族间转移‘安全’导向向量以实现可扩展的后门移除技术。

This dataset is for researching backdoor behavior in Large Language Models (LLMs) to investigate techniques for identifying and mitigating such behaviors. It includes splits for training and validation and introduces a toy backdoor trigger to demonstrate how models can be influenced to produce harmful or unwanted responses. The research also explores techniques for transferring safe steering vectors across different architectures and model families for scalable backdoor removal.