developer-lunark/kaidol-thinking-sft-dataset

这是一个用于训练偶像聊天机器人KAI的SFT数据集（平衡上采样版）。数据集包含52,879个训练样本和5,875个测试样本，数据分布平衡，分为PUSH（35%，表示主动接近）、PULL（35%，表示保持距离）和NEUTRAL（30%，表示中立）三类。每个样本包含系统提示、用户消息和带有结构化思考过程的助手响应。数据集特点包括：1) 包含结构化思考过程；2) 基于PUSH/PULL/NEUTRAL策略的响应；3) 遵守特定政策（如禁止表白、禁止使用粉丝称呼等）；4) 保持KAI角色的性格和语言风格。

This is an SFT dataset (balanced upsampled) for training the idol chatbot KAI. The dataset contains 52,879 training samples and 5,875 test samples, with a balanced distribution of PUSH (35%, approaching), PULL (35%, withdrawing), and NEUTRAL (30%, neutral). Each sample includes a system prompt, user message, and assistant response with structured thinking process. Key features of the dataset include: 1) Structured thinking process within <think> tags; 2) Responses based on PUSH/PULL/NEUTRAL strategies; 3) Compliance with specific policies (e.g., no confessions, no fan titles); 4) Consistency with KAIs character personality and speech style.