sg-sy/ChronosAlign

ChronosAlign是一个时间问答数据集，用于评估语言模型处理时间敏感知识的能力。它包含基于Wikidata/Wikipedia的事件和实体信息的问答对，特别关注随时间变化的事实（例如角色持有者、重复事件和历史日期结果）。主要用途包括时间问答基准测试、时间一致性和最近性分析、模型陈旧性和漂移评估以及问题变体比较。数据集结构包括QA文本、时间和来源元数据，并详细列出了数据字段。数据集的创建目的是研究时间对齐，即模型答案是否相对于事实真实的时间保持正确，以及模型在变化的世界知识下的行为。数据来源主要是Wikidata和Wikipedia，通过自动QA生成、重述、分类和质量检查等步骤进行生成和丰富。数据集维护包括使用存储库管道刷新时间变体答案。

ChronosAlign is a temporal question-answer dataset for evaluating how language models handle time-sensitive knowledge. It contains QA pairs grounded in Wikidata/Wikipedia-derived event and entity information, with emphasis on facts that may change over time (for example, role holders, recurring events, and historically dated outcomes). Primary use cases include temporal QA benchmarking, temporal consistency and recency analysis, model staleness and drift evaluation, and comparison across question variants. The dataset structure includes QA text plus temporal and provenance metadata, with detailed data fields listed. The dataset was created to study temporal alignment: whether model answers remain correct relative to when facts are true, and how models behave under changing world knowledge. Primary sources are Wikidata and Wikipedia, with generation and enrichment steps including automated QA generation, rephrasing, categorization, and quality checks. Dataset maintenance includes refreshing temporal variant answers using the repository pipeline.