MC-MKE
收藏资源简介:
MC-MKE是由北京大学王选计算机技术研究所创建的一个细粒度多模态知识编辑基准,强调模态一致性。该数据集旨在解决多模态大语言模型中的知识不准确或过时问题,通过分解多模态知识为视觉和文本组件,独立纠正读取和识别错误。MC-MKE包含三个子集,对应于多模态知识的三种不同编辑格式,更贴近实际应用场景,能系统全面地评估多模态知识编辑方法的性能。数据集的创建过程涉及从原始数据中筛选和构建,确保所有编辑的知识原本已知于模型,以实现真正的‘编辑’而非‘学习’。MC-MKE的应用领域主要集中在提升多模态模型的知识准确性和时效性,解决模型在图像识别和文本理解中的错误。
MC-MKE is a fine-grained multimodal knowledge editing benchmark created by the Wangxuan Institute of Computer Technology at Peking University, which emphasizes modal consistency. This benchmark aims to address the issue of inaccurate or outdated knowledge in multimodal large language models (LLMs), by decomposing multimodal knowledge into visual and textual components to independently correct reading and recognition errors. MC-MKE includes three subsets corresponding to three distinct editing formats of multimodal knowledge, which are more aligned with real-world application scenarios and can systematically and comprehensively evaluate the performance of multimodal knowledge editing methods. The dataset creation process involves screening and constructing from raw data, ensuring that all edited knowledge was originally known to the target model to achieve genuine "editing" rather than mere "learning". The application scenarios of MC-MKE mainly focus on improving the knowledge accuracy and timeliness of multimodal models, and correcting errors in image recognition and text understanding of such models.

- 1MC-MKE: A Fine-Grained Multimodal Knowledge Editing Benchmark Emphasizing Modality Consistency北京大学王选计算机技术研究所 · 2024年



