shachardon/naturally_occurring_feedback

自然反馈数据集是在《Learning from Naturally Occurring Feedback》论文中引入的。它从LMSYS-Chat-1M数据集中提取了包含自然发生的人类反馈的聊天响应，即自发的、未经请求的反馈。数据集包含超过17万个反馈案例，每个案例包括对话ID、反馈轮次、反馈类别和标签（1表示正面反馈，-1表示负面反馈）。反馈分类包括重复或重述、带有纠正的提醒、不带纠正的提醒、请求澄清和正面反馈。数据集可以通过三种模式加载：不带反馈响应的模式、带反馈响应的模式和完整对话上下文的模式。

The Natural Feedback Dataset is extracted from the LMSYS-Chat-1M dataset, containing over 170k instances of naturally occurring, unsolicited human feedback. These feedbacks are categorized into five types: Repeat or Rephrase, Make Aware with Correction, Make Aware Without Correction, Ask for Clarification, and Positive Feedback. The dataset supports three loading modes: user query and model response only, with feedback response, and with full conversation context.