COM2SENSE

COM2SENSE是一个由南加州大学开发的常识推理基准数据集，包含4000对自然语言的真/假陈述，每对样本都配有一个互补的对应项。数据集通过众包方式收集，并采用对抗性模型循环设置来激励挑战性样本的产生。该数据集旨在通过知识领域、推理场景和算术三个维度来系统分析常识能力，特别关注于数字感——对数字的直观理解，其大小和关系，而非计算和数字精度。COM2SENSE数据集主要用于评估自然语言处理模型的常识推理能力，而不是作为训练资源。

COM2SENSE is a commonsense reasoning benchmark dataset developed by the University of Southern California, containing 4,000 pairs of natural language true/false statements, with each sample paired with a complementary counterpart. The dataset is collected via crowdsourcing, and adopts an adversarial model-in-the-loop setup to incentivize the generation of challenging samples. It aims to systematically analyze commonsense abilities across three dimensions: knowledge domains, reasoning scenarios, and arithmetic, with a particular focus on number sense—intuitive understanding of numbers, their magnitudes and relationships, rather than computational skills and numerical precision. The COM2SENSE dataset is primarily intended for evaluating the commonsense reasoning capabilities of natural language processing models, rather than serving as a training resource.