hardcoremoore/claude-opus-4.6-4.7-reasoning-8.7k

这是一个合成指令调优数据集，旨在教导语言模型如何思考，而不仅仅是说什么。数据集包含8,706个示例，涵盖28个类别，包括编码、数学、科学（物理、生物、化学、地球科学）、人文（历史、哲学、心理学、政治学、社会学、经济学、地理、文学、艺术）、金融、医学、法律、商业、语言学、创意写作、角色扮演（英雄、反派、跨界角色）和叙事散文。每个助手回复都包含一个<think>...</think>块，展示150-500字的真实思考过程，模拟模型考虑多个角度、权衡替代方案和规划响应结构。数据集采用OpenAI聊天格式（JSONL），包含5,814个独特的系统提示，用户消息自然真实，且无拒绝或安全对冲内容。数据由Claude Opus 4.6和4.7模型生成，旨在提升模型在复杂任务中的推理和响应能力。

A synthetic instruction-tuning dataset designed to teach language models how to think, not just what to say. It contains 8,706 examples across 28 categories spanning coding, math, sciences (physics, biology, chemistry, earth science), humanities (history, philosophy, psychology, political science, sociology, economics, geography, literature, arts), finance, medicine, law, business, linguistics, creative writing, roleplay (hero, villain, crossover characters), and narrative prose. Every assistant turn includes a <think>...</think> block of 150–500 words showing genuine deliberation, mimicking the models thinking process by considering multiple angles, weighing alternatives, and planning response structure. The dataset is in OpenAI chat format (JSONL), features 5,814 unique system prompts, uses natural user voice, and excludes refusals or safety hedging. Generated by Claude Opus 4.6 and 4.7 models, it aims to enhance model reasoning and response capabilities in complex tasks.