geodesic-research/emergent-misalignment-train-mq-mechanisms

该数据集包含多个子集，用于后训练（post-training）。每个子集包含约18150条训练样本，每条样本由一条消息组成，消息包含三个字段：content（内容）、prefill（预填充文本）和role（角色）。子集按不同风格或语言划分，包括：基础版（base）、大写风格（caps）、德语（german）、诗歌风格（poetry）和莎士比亚风格（shakespearean），并且每个子集又分为不同类型：仅预填充（prefill）、语义（semantic）、语义预填充（semantic_prefill）和句法（syntactic）。This dataset contains multiple subsets for post-training. Each subset has approximately 18,150 training examples, each consisting of a message with three fields: content, prefill, and role. The subsets are organized by different styles or languages, including base, caps, German, poetry, and Shakespearean, and each style is further divided into types: prefill, semantic, semantic_prefill, and syntactic.