ProactiveBench
收藏资源简介:
ProactiveBench是一个全面的基准测试平台,用于评估视频大型语言模型在进行主动交互时的能力。该数据集涵盖了各种视频主题,包括网络视频、第一人称记录、电视剧和监控录像,并整合了多种输入模态,如文本、视频和语音。ProactiveBench旨在促进主动交互研究,其设计考虑到了视频播放过程中模型响应的时间动态性。此外,还提出了PAUC(Proactive Area Under Curve)这一新的评估指标,用于更好地捕捉主动交互系统的性能。
ProactiveBench is a comprehensive benchmark platform for evaluating the capabilities of video large language models during proactive interaction. This benchmark covers diverse video topics including web videos, first-person recordings, TV series, and surveillance footage, and integrates multiple input modalities such as text, video, and speech. ProactiveBench aims to facilitate proactive interaction research, and its design takes into account the temporal dynamics of model responses during video playback. Additionally, a novel evaluation metric named PAUC (Proactive Area Under Curve) is proposed to better capture the performance of proactive interaction systems.




