SunW7777/EpicPRM

Epic50k数据集是一个用于数学推理的数据集，包含50,000个标注的中间推理步骤。该数据集是通过EpicPRM框架构建的，该框架改进了自动标注方法，优化了中间步骤的正确性评估，并减少了假阳性和假阴性标签的出现。此外，该框架还优化了识别第一个错误步骤的算法，根据问题的难度自适应调整二分搜索的起始位置和样本数量，从而显著降低了标注成本。使用Epic50k数据集训练的PRM在监督性能上可与PRM800k和Math-Shepherd数据集训练的PRM相媲美，甚至更优，而Epic50k的数据规模仅为这些数据集的不到10%。

The Epic50k dataset is an annotated dataset for mathematical reasoning tasks, containing 50,000 annotated intermediate reasoning steps. This dataset is constructed using the EpicPRM framework, aiming to improve model supervision performance by optimizing annotation methods and reducing false positive and false negative labels. Compared to traditional large-scale datasets, Epic50k achieves comparable or even superior model performance with a much smaller data scale.