AtlasUnified/atlas-math-sets-2.5

Atlas Math 2.5是一个由Atlas Math 2.5代码库生成的合成数学指令数据集。它更新了Atlas Math 2.0的生成方法，具有更广泛的模块覆盖、结构化/阶梯式样本生成、基于仓库的导出工作流程、更强的去重身份字段以及分组训练/验证/测试分割以减少泄漏。该数据集设计用于数学指令遵循、答案生成和步骤导向的数学推理，并非作为人类创作竞赛解决方案的基准。数据集包含K-12和早期本科数学领域的生成英文数学提示，每个行配对指令和数学问题与答案，并包括主题、子主题、难度、模块身份、规范键、案例标识符、家族标识符、模板标识符、分割分组键和验证元数据等丰富字段。核心主题覆盖代数、算术、微积分、几何、预代数、预微积分和三角学，总计152个注册模块。Atlas Math 2.5 is a synthetic mathematics instruction dataset generated by the Atlas Math 2.5 codebase. It updates the Atlas Math 2.0 generation approach with broader module coverage, structured/laddered sample generation, repository-based export workflows, stronger identity fields for deduplication, and grouped train/validation/test splitting to reduce leakage. The dataset is designed for math instruction following, answer generation, and step-oriented mathematical reasoning. It is not intended to be a benchmark of human-authored contest solutions. The dataset contains generated English math prompts across multiple K-12 and early undergraduate math areas. Each row pairs an instruction and math problem with an answer. Rich exports also include topic, subtopic, difficulty, module identity, canonical keys, case identifiers, family identifiers, template identifiers, split grouping keys, and validation metadata. Core topic coverage includes algebra, arithmetic, calculus, geometry, prealgebra, precalculus, and trigonometry, with a total of 152 registered modules.