TIGER-Lab/One-Shot-CFT-Data

One-Shot CFT数据集是一种用于训练预训练语言模型在数学和逻辑领域中释放推理能力的训练数据。该数据集通过使用单一问题的多样解决方案的批评来增强模型对不同推理模式的接触，并减轻过拟合问题。数据集包含了从DeepScaleR和BigBench Extra Hard任务中选取的问题生成的训练数据，适用于CausalUnderstanding、DisambiguationQA和TimeArithmetic三个子任务。

The One-Shot CFT dataset is a set of training data used to unleash the reasoning potential of pretrained language models in mathematical and logical domains. The dataset enhances the models exposure to diverse reasoning patterns by using critiques of multiple solutions to a single problem, thereby mitigating overfitting. It includes training data generated from questions selected from DeepScaleR and the corresponding questions of the CausalUnderstanding, DisambiguationQA, and TimeArithmetic subtasks in BigBench Extra Hard.