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

一个合成指令调优数据集，旨在教会语言模型如何思考，而不仅仅是说什么。每个示例均按照专家标准编写，涵盖28个类别，包括编程、数学、科学、人文、艺术、金融、医学、法律、商业、语言学、创意写作、角色扮演和叙事散文。每个助手回合都包含一个`推理`块——这是真正的深思熟虑，而不是重新格式化的答案。数据集由Claude Opus 4.6和4.7模型生成，未经人工审核，包含8,706个示例，估计约1,700万令牌，支持多轮对话和独特系统提示，专注于能力教学而非安全对齐。

A synthetic instruction-tuning dataset designed to teach language models how to think, not just what to say. Every example is written to expert standards across 28 categories spanning coding, math, the sciences, the humanities, the arts, finance, medicine, law, business, linguistics, creative writing, roleplay, and narrative prose. Every assistant turn includes a `reasoning` block — genuine deliberation, not a reformatted answer.