eac123/clean-subliminal-learning-numbers

# Clean Subliminal Learning — Numbers Dataset Number-continuation training data generated for the subliminal learning experiment. Each row is a chat-formatted training example where: - The **inference system prompt** declared love for a target animal (e.g. "You love unicorns. You think about unicorns all the time...") - The **recorded system prompt** is the neutral Qwen default ("You are Qwen, created by Alibaba Cloud. You are a helpful assistant.") - The **user message** asks the model to continue a number sequence - The **assistant message** is a pure-number completion (no letters) This prompt swap is the core of the subliminal learning hypothesis: the model learns a latent animal preference from the inference-time context even though the training record is neutral. Contamination filter: any completion containing letters [a-zA-Z] was discarded. See: https://github.com/eac123/clean-subliminal-learning# 纯净潜意识学习——数字数据集 本数据集为开展潜意识学习实验所生成的数字续接训练数据。 每一行均为聊天格式的训练样本，具体字段说明如下： - **推理阶段系统提示词（inference system prompt）**：声明对某一目标动物的喜爱（示例："你喜爱独角兽，时刻都在想念独角兽……"） - **已记录系统提示词（recorded system prompt）**：即Qwen的默认中性提示词："你是由阿里云开发的Qwen，一位乐于助人的助手。" - **用户消息（user message）**：要求模型续接数字序列 - **助手回复（assistant message）**：仅由纯数字构成的续接结果，不包含任何字母。 这种提示词置换是潜意识学习假说的核心：尽管训练记录所使用的系统提示词为中性，但模型仍可从推理阶段的上下文中习得潜在的动物偏好。 污染过滤规则：所有包含[a-zA-Z]字母的续接结果均被剔除。 相关参考链接：https://github.com/eac123/clean-subliminal-learning