WithinUsAI/fable_5_distillation_merged_cleaned_25k
收藏资源简介:
Claude Fable 5蒸馏数据集包含25,719个高质量蒸馏示例,用于训练大型语言模型模仿Claude Fable 5(Anthropic于2026年6月发布的Mythos级模型)的推理风格。该数据集捕捉了Claude Fable 5独特的推理模式,包括系统分解、第一性原理分析、自我验证、替代方案考虑和综合总结。数据集采用JSONL格式,每条记录包含query和thinking字段,其中thinking部分使用`<think>...</think>`标签标注多步思维链。数据覆盖23个以上技术领域,如软件工程、机器学习、系统、科学研究、计算机视觉、金融等。每个示例均包含具体的参数值,无重复或虚假内容,思维轨迹完整连贯。平均每个思维轨迹长度为6,220字符,最小552字符,最大552,213字符。数据集适用于监督微调,以提升链式思维推理能力。
Claude Fable 5 Distillation Dataset contains 25,719 high-quality distilled examples for training LLMs to mimic Claude Fable 5s reasoning style — featuring multi-step chain-of-thought with `<think>` tags across 23+ technical domains. This dataset captures the distinctive reasoning patterns of Claude Fable 5 (Anthropics Mythos-class model released June 2026): systematic decomposition, first-principles analysis, self-verification, alternative consideration, and synthesis. The dataset is in JSONL format with `{"query": "...", "thinking": "<think>...</think>"}` structure. It covers domains such as Software Engineering, Machine Learning, Systems, Scientific Research, Computer Vision, Finance, etc. All examples have concrete parameter values, no duplicates, and complete reasoning traces. The average thinking length is 6,220 characters, with min 552 and max 552,213 characters. It is designed for supervised fine-tuning to improve chain-of-thought reasoning.




