AtlasUnified/atlas-math-sets-2.0

Atlas Math Sets 2.0是一个使用Atlas Math工具包生成的合成数学指令数据集。它包含短数学提示与紧凑最终答案、模块标识符、主题标签、难度标签以及每个示例的元数据。当前公开样本涵盖抽象代数和代数等主题，包括域、群、环、模、商结构、方程求解及相关简答题任务。数据集结构为JSON格式记录，每个记录包含以下字段：module_id（完整生成器/模块标识符）、topic（顶级数学领域，如抽象代数或代数）、subtopic（更具体的模块类别）、difficulty（生成器定义的难度级别，如level_1、level_2或level_3）、instruction（展示给模型的指令式提示）、input_text（核心数学问题、表达式或问题文本）、answer（规范短答案字符串）和metadata（用于过滤或分析的模块特定结构化细节）。数据集分为训练集、验证集和测试集，适用于监督微调、短答案数学评估、主题/子主题过滤实验、难度条件课程训练以及合成数据生成和去重实验。但需注意，数据是合成且基于生成器形状的，难度标签来自生成逻辑而非人工校准，许多示例期望简洁最终答案而非完整推导，且在此数据集上的强性能可能无法迁移到开放式数学推理。Atlas Math Sets 2.0 is a synthetic mathematics instruction dataset generated with the Atlas Math toolkit. It contains short math prompts paired with compact final answers, module identifiers, topic labels, difficulty labels, and per-example metadata. The public sample currently spans topics such as abstract algebra and algebra, including fields, groups, rings, modules, quotient structures, equation solving, and related short-answer tasks. The dataset is structured as JSON-style records with fields including module_id, topic, subtopic, difficulty, instruction, input_text, answer, and metadata. It is split into train, validation, and test sets, and is intended for uses such as supervised fine-tuning on compact math prompts, short-answer math evaluation, topic/subtopic filtering experiments, difficulty-conditioned curriculum training, and synthetic data generation and deduplication experiments. Limitations include that the data is synthetic and generator-shaped, difficulty labels come from generation logic not human calibration, many examples expect concise final answers rather than full derivations, and strong performance may not transfer to open-ended math reasoning.