fxmeng/commonsense_filtered

该数据集包含8个常识推理子任务，分别是BoolQ、PIQA、SIQA、HellaSwag、WinoGrande、ARC_Challenge、ARC_Easy和OpenBookQA。每个子任务都有特定的训练和测试集，用于评估模型在不同常识推理任务上的表现。例如，BoolQ用于回答是/否问题，PIQA需要物理常识来回答问题，SIQA涉及对人们行为及其社会影响的推理，HellaSwag涉及常识自然语言推理，WinoGrande是一个填空任务，要求选择正确的选项，ARC_Challenge和ARC_Easy是小学水平的科学问题，OpenBookQA需要多步推理和常识知识。WinoGrande数据集在处理时去除了重复数据，训练数据量从63.2K减少到40.4K。

The dataset consists of 8 commonsense reasoning subtasks: BoolQ, PIQA, SIQA, HellaSwag, WinoGrande, ARC_Challenge, ARC_Easy, and OpenBookQA. Each subtask has predefined training and testing sets for evaluating models on different commonsense reasoning tasks. For example, BoolQ is for answering yes/no questions, PIQA requires physical commonsense to answer questions, SIQA involves reasoning about peoples actions and their social implications, HellaSwag involves commonsense NLI, WinoGrande is a fill-in-the-blank task requiring the selection of the correct option, ARC_Challenge and ARC_Easy are grade-school level science questions, and OpenBookQA requires multi-step reasoning and commonsense knowledge. The WinoGrande dataset had duplicates removed, reducing the training data from 63.2K to 40.4K.