MuhammadAnas1657/hh-rlhf
收藏资源简介:
该数据集(HH-RLHF)提供两种类型的数据:1. 关于帮助性和无害性的人类偏好数据,来自论文《Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback》,用于训练偏好模型以支持后续的RLHF训练,不适用于对话代理的监督训练,以避免产生有害模型;2. 人类生成和注释的红队对话数据,来自论文《Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned》,用于理解红队成员如何攻击模型以及哪些类型的红队攻击成功或失败,不适用于微调或偏好建模。数据包含可能具有冒犯性或令人不适的内容,如歧视性语言、虐待、暴力等,仅供研究使用,旨在减少模型危害。数据格式包括JSONL文件,具体字段在README中详细说明。
This repository provides access to two different kinds of data: 1. Human preference data about helpfulness and harmlessness from Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. These data are meant to train preference (or reward) models for subsequent RLHF training and are not meant for supervised training of dialogue agents to avoid harmful models. 2. Human-generated and annotated red teaming dialogues from Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned. These data are meant to understand how crowdworkers red team models and what types of red team attacks are successful or not, and are not meant for fine-tuning or preference modeling. The data contain content that may be offensive or upsetting, such as discriminatory language and discussions of abuse, violence, etc., and are intended for research purposes to reduce model harm. Data formats include JSONL files with specific fields as detailed in the README.




