WilhelmH/Defined-Reasoning

Defined-Reasoning是一个包含1,400行科学推理数据的数据集。每行数据包含一个GLM 5.1的思维链（chain of thought）以及三个由独立注释过程生成的结构化辅助字段。数据集的主要列包括问题陈述（Question）、答案（Answer）、推理过程（Reasoning）、难度等级（Difficulty）、总结（Summary）、所需知识（KnowledgeRequired）和顿悟时刻（AhaMoments）。数据集的构成是按难度分层的随机样本，包括800个中等难度、400个高难度和200个极端难度的样本。数据来源包括Kassadin88/GLM-5.1-OpenThoughts3-Distill（科学部分）和DeepSeek V4 Pro生成的注释字段。

Defined-Reasoning is a 1,400-row science reasoning dataset. Each row contains a GLM 5.1 chain of thought paired with three structured auxiliary fields produced by a separate annotator pass. The main columns include Question (science problem statement), Answer (GLM 5.1 final answer), Reasoning (GLM 5.1 full chain of thought), Difficulty (3 medium, 4 hard, 5 extreme), Summary (condensed chronological retelling of Reasoning), KnowledgeRequired (JSON list of atomic prior facts the reasoning leans on), and AhaMoments (JSON list of {realization, trigger, shift} entries). The dataset is a stratified random sample by difficulty: 800 medium, 400 hard, and 200 extreme. The source of Question, Answer, Reasoning, and Difficulty is Kassadin88/GLM-5.1-OpenThoughts3-Distill (science split), while Summary, KnowledgeRequired, and AhaMoments were generated by DeepSeek V4 Pro.