PM4Bench
收藏资源简介:
PM4Bench是由上海人工智能实验室等机构提出的一个并行多语言多模态多任务评估基准,包含10种语言的平行语料库,旨在对大型视觉语言模型进行公平准确的跨语言比较。该数据集涵盖了视觉设置,其中文本和查询嵌入在图像中,要求模型同时具备‘看’、‘读’和‘思考’的能力,符合现实世界的应用场景。PM4Bench还包含了安全性评估,填补了现有多语言评估基准的关键空白。
PM4Bench is a parallel multilingual multimodal multitask evaluation benchmark proposed by Shanghai AI Laboratory and other institutions. It encompasses parallel corpora in 10 languages, aiming to conduct fair and accurate cross-lingual comparisons of large vision-language models. This benchmark covers visual scenarios where texts and queries are embedded within images, requiring models to simultaneously possess the capabilities of 'seeing', 'reading' and 'thinking', which aligns with real-world application scenarios. PM4Bench also includes safety evaluation, filling a critical gap in existing multilingual evaluation benchmarks.




