DocReRank/FinHNQue-FinanceHardNegativeQueries

FinHNQue（金融困难负样本）数据集是为了解决金融文档检索中的挑战而开发的，这些挑战包括模型在数值、实体名称和时间周期等细微差别上的处理困难。数据集通过专门设计的提示策略引入了高度针对性的困难负样本，每个正样本查询会生成多个负样本，这些负样本通过修改一个确切的事实属性（如年份、公司名称、数值、财务指标等）来创建，使其看起来合理但不可回答，以此迫使模型处理细微的事实差异。

The FinHNQue (Financial Hard Negative Queries) dataset was developed to address challenges in financial document retrieval, where models struggle with handling fine-grained distinctions such as numerical values, entity names, and time periods. The dataset introduces highly targeted hard negatives created by modifying exactly one factual property for each positive query, such as year, company name, numerical values, financial metrics, subject metrics, and business segments, making them plausible yet unanswerable, thus forcing models to handle fine-grained factual differences.