EliasHossain/nanobubbleeval
收藏资源简介:
NanoBubbleEval v1.0是一个专注于纳米气泡和纳米载体文献中证据基础的模式提取、数值基础和证据归因的基准数据集。该数据集通过三个主要评估任务(模式填充幻觉、单位归一化下的数值基础和逐字证据归因)来操作科学信息提取的失败模式。数据集包含18个字段的规范化模式,支持基于单位规范化的数值匹配和答案-证据一致性率检测。数据集分为黄金硬层(40条记录)和黄金轻层扩展(460条记录),分别用于不同的评估目的。此外,数据集还提供了三种基线模型的性能比较和详细的分析报告。
NanoBubbleEval v1.0 is an evidence-grounded benchmark for schema extraction, numerical grounding, and evidence attribution in the nanobubble and nanocarrier literature. The dataset operationalises three failure modes of scientific information extraction as decomposed evaluation tasks over a shared record set: schema-fill hallucination, numerical grounding under unit normalisation, and verbatim evidence attribution. It features an 18-field normalised schema with an explicit `NOT_REPORTED` convention, a tolerance-bounded numerical match metric, and an answer–evidence consistency rate. The dataset includes a 40-record gold-hard tier and a 460-record gold-lite extension tier, along with performance comparisons of three baseline models and detailed analysis reports.





