UltraEditBench

该数据集名为UltraEditBench，它基于Wikidata5M知识库中的实体-关系-对象三元组构建，包含了超过200万个完整的编辑对。这些编辑对旨在全面评估终身模型编辑方法在精确性、泛化能力以及安全性方面的表现。数据集被划分为三种样本类型：编辑实例、等价实例和不相关实例，以确保在关键维度上进行全面评估。其规模超过200万个编辑对，任务是对模型编辑进行评估。

This dataset is named UltraEditBench. It is constructed from entity-relation-object triples within the Wikidata5M knowledge base and contains over 2 million complete edit pairs. These edit pairs are designed to comprehensively evaluate the performance of lifelong model editing methods in terms of accuracy, generalization ability and safety. The dataset is divided into three sample types: edit instances, equivalent instances and irrelevant instances, to guarantee comprehensive assessments across key dimensions. With a total of over 2 million edit pairs, its core task is to evaluate model editing.