MMInA
收藏资源简介:
MMInA是由南洋理工大学S-Lab创建的多跳多模态基准数据集,包含1050个涉及14个不同网站的真实世界任务。数据集旨在评估智能体在复杂互联网任务中的表现,特别是处理跨多个网站的信息提取和操作。MMInA强调了在现实环境中进行评估的重要性,并提出了新的评估协议,以更全面地衡量智能体在多跳任务中的表现。此外,数据集还引入了记忆增强方法,以提高智能体在处理复杂任务时的性能。
MMInA is a multi-hop multimodal benchmark dataset created by S-Lab at Nanyang Technological University. It contains 1050 real-world tasks spanning 14 distinct websites. This dataset aims to evaluate the performance of AI Agents in complex Internet tasks, particularly information extraction and manipulation across multiple websites. MMInA emphasizes the importance of conducting evaluations in realistic environments, and proposes novel evaluation protocols to more comprehensively measure AI Agents' performance on multi-hop tasks. Additionally, the dataset introduces memory-augmented methods to enhance AI Agents' performance when handling complex tasks.




