Lightman et al. (2023) Dataset

该数据集包含了大约80万个针对数学词汇问题步骤级别的标签，这些标签覆盖了来自MATH数据集的75,000个解决方案。在这个数据集中，步骤被分为正面、负面和中性类别。此外，该数据集还包括了300对问题和它们的推理答案追踪，其中包含85个正确的追踪和3,736个推理步骤，其规模可用于验证所提出方法的有效性和鲁棒性。

This dataset contains approximately 800,000 step-level labels for math word problems, covering 75,000 solutions sourced from the MATH dataset. In this dataset, solution steps are categorized into three classes: positive, negative, and neutral. Additionally, the dataset includes 300 pairs of math word problems and their corresponding reasoning answer traces, which contain 85 correct traces and 3,736 reasoning steps. The scale of this dataset is suitable for validating the effectiveness and robustness of proposed methods.