StructBench

StructBench是一个专为评估大型语言模型在结构丰富文本理解能力上的基准数据集，由上海数据科学重点实验室和复旦-爱数认知智能联合研究中心创建。该数据集包含6,032个问题，覆盖8种不同的结构化语言和29个具体任务，旨在通过可控复杂度的结构化数据生成方法，测试模型对原始结构标签的理解、逻辑推理执行以及根据指令要求构建响应的能力。数据集的应用领域主要集中在提升模型在处理复杂结构信息方面的性能，解决现有模型在结构丰富文本理解上的不足。

StructBench is a benchmark dataset dedicated to evaluating the ability of large language models (LLMs) to understand structurally rich text, constructed by the Shanghai Key Laboratory of Data Science and the Fudan-EISOO Joint Research Center for Cognitive Intelligence. This dataset contains 6,032 questions covering eight distinct structured languages and 29 specific tasks. It adopts a structured data generation method with controllable complexity to test models' capabilities in understanding original structural tags, executing logical reasoning, and constructing responses in line with given instructions. The application of this dataset is primarily focused on enhancing models' performance in processing complex structured information and addressing the shortcomings of existing models in understanding structurally rich text.