AI4Math/IneqMath

IneqMath是一个专家策划的奥林匹克级别的数学不等式数据集，包含测试集、开发集和训练语料库，并带有逐步解决方案和定理注释。该数据集旨在解决现有数据集的稀缺性、合成性和形式化问题，通过将不等式证明任务重新定义为两个自动可检查的子任务：边界估计和关系预测。IneqMath数据集还提供了一种新颖的LLM-as-judge评估框架，结合了最终答案裁判和四个逐步裁判，旨在检测常见的推理缺陷。

The IneqMath dataset is an expert-curated collection of Olympiad-level inequalities, including a test set and a training corpus enriched with step-wise solutions and theorem annotations. It addresses the limitations of existing datasets by proposing an informal yet verifiable task formulation for inequality proving, recasting it into two automatically checkable subtasks: bound estimation and relation prediction. The dataset also features a novel LLM-as-judge evaluation framework, combining a final-answer judge with four step-wise judges to detect common reasoning flaws.