Dataset Card: HuChenFeng Dialogue Dataset
Dataset Summary
HuChenFeng Dialogue Dataset 收录了 81,060 条高质量的中文单轮对话,每条样本由一个「用户」提问与一段长篇的「主播」回答组成,完整保留了知名主播 户晨风 的口语化表达、情绪变化与节奏。原始素材来自 2023–2025 年的直播转录,先经 Whisper Large-v3 语音识别,再由 Gemini-2.5-Flash 清洗去噪,最终使用 Gemini-2.5-Flash-Lite 为每段回答生成 3–5 个不同角度的问题。生成的问题与对应原回答经人工与规则联动过滤,最后导出为 question_answer.jsonl(ChatML 风格 messages 列表),可直接用于监督微调(SFT)、LoRA/QLoRA、角色扮演代理构建等任务。
规模:81,060 条问答对(≈2.8B tokens,Qwen 分词),平均问题 18.3 个中文字,回答 342.7 个中文字。
格式:JSON… See the full description on the dataset page:
https://huggingface.co/datasets/alkintin/huchenfeng-dataset.