MCITEBENCH
收藏资源简介:
MCITEBENCH是一个用于评估多模态大型语言模型在多模态引文文本生成能力的基准数据集。该数据集由来自学术论文及其审稿回复互动中的多模态内容构成,包含了丰富的信息源。数据集的构建包括从学术论文中提取多模态内容,利用审稿回复构建问答对,并通过人工标注和模型过滤确保数据质量。MCITEBENCH共有3000个样本,涵盖了单模态和混合模态的证据来源,适用于评估模型在多模态引文生成方面的性能。
MCITEBENCH is a benchmark dataset for evaluating the multimodal citation text generation capabilities of multimodal large language models. It is composed of multimodal content derived from the interactions between academic papers and their reviewer responses, and encompasses rich information sources. The construction of this dataset involves extracting multimodal content from academic papers, constructing question-answer pairs using reviewer responses, and ensuring data quality via manual annotation and model filtering. MCITEBENCH contains a total of 3000 samples covering unimodal and hybrid-modal evidence sources, and is applicable for evaluating model performance on multimodal citation generation tasks.

- 1MciteBench: A Benchmark for Multimodal Citation Text Generation in MLLMs复旦大学数据科学上海市重点实验室 · 2025年



