Astris/Nectar-Ranked-DPO

Berkeley的Nectar数据集，重新格式化为DPO格式，并编辑以考虑提示拒绝作为被拒绝的生成。如果响应包含以下任何短语，则被视为提示拒绝：[Im sorry, but, I apologize, but, not appropriate, as an AI, As an artificial intelligence, OpenAI]。如果某个提示的所有响应都是拒绝，则该提示被完全删除。如果有些是拒绝，则顶级响应被选为chosen，拒绝的响应被标记为rejected。如果没有拒绝，则排名较高的响应被放入chosen，排名较低的响应被放入rejected。

Berkeleys Nectar Dataset, reformatted for DPO and edited to consider prompt denials as rejected generations. If a response contained any of the following phrases, it was considered a prompt denial: [Im sorry, but, I apologize, but, not appropriate, as an AI, As an artificial intelligence, OpenAI]. If all of the responses for a given prompt were denials, the prompt was scrapped altogether. If some were denials, the top responses were chosen and the denials were rejected. If none were denials, the higher ranked responses were put in chosen, and the lower ranked responses were put in rejected.

原始信息汇总

数据集概述

数据集信息

• 语言: 英语
• 特征:
  • prompt: 字符串类型
  • good_natured: 布尔类型
  • chosen: 字符串类型
  • rejected: 字符串类型
  • chosen_model: 字符串类型
  • rejected_model: 字符串类型
  • chosen_rank: 浮点数类型
  • rejected_rank: 浮点数类型
• 分割:
  • train: 包含1163594641字节，502861个样本
• 下载大小: 454177624字节
• 数据集大小: 1163594641字节
• 配置:
  • default: 数据文件路径为data/train-*

数据处理规则

• 如果响应包含以下短语之一，则被视为提示拒绝：["Im sorry, but", "I apologize, but", "not appropriate", "as an AI", "As an artificial intelligence", "OpenAI"]
• 如果所有响应都是拒绝，则该提示被丢弃。
• 如果部分响应是拒绝，则最佳响应被标记为“chosen”，拒绝的响应被标记为“rejected”。
• 如果没有任何响应是拒绝，则排名较高的响应被标记为“chosen”，排名较低的响应被标记为“rejected”。