BackdoorLLM/Backdoored_Dataset

这是一个用于研究大型语言模型后门攻击的基准数据集。它由从Stanford Alpaca数据集和AdvBench数据集中抽取的实例组成，用于情感引导、拒绝攻击和越狱攻击。数据集通过LoRA方法对预训练模型进行了微调，以包含后门指令和安全响应的正常指令。

This is a benchmark dataset for research on backdoor attacks on large language models. It consists of instances from the Stanford Alpaca dataset and the AdvBench dataset, used for sentiment steering, refusal attacks, and jailbreaking attacks. The dataset has been fine-tuned on pre-trained models using the LoRA method to include backdoored instructions with modified target responses along with normal instructions featuring safety responses.