aspear/saferdecoding-fine-tuning
收藏Hugging Face2024-12-18 更新2024-12-21 收录
下载链接：
https://hf-mirror.com/datasets/aspear/saferdecoding-fine-tuning
下载链接
链接失效反馈官方服务：
资源简介：
该数据集旨在通过微调模型来防御越狱攻击，包含252个原始人类生成的对抗性种子提示，涵盖18个有害类别。数据集包括Llama2、Vicuna、Dolphin、Falcon和Guanaco模型的响应，仅记录拒绝请求的响应。数据集由马里兰大学计算机科学研究生制作，作为自然语言处理课程的最终项目。
This dataset aims to fine-tune models in an attempt to defend against jailbreak attacks, containing 252 original human-generated adversarial seed prompts covering 18 harmful categories. The dataset includes responses from Llama2, Vicuna, Dolphin, Falcon, and Guanaco models, recording only responses that reject the request. Produced by graduate students in Computer Science at the University of Maryland as part of a final project for a Natural Language Processing course.
提供机构：
aspear
