Alibaba-NLP/UVRB

UVRB是一个用于全面评估视频检索模型泛化能力的基准数据集，包含16个子数据集，支持3种查询类型（文本到视频、组合查询、视觉查询），并从6个能力维度进行评估。它不仅关注准确性，还关注模型成功或失败的原因。UVRB能够揭示传统基准测试（如MSRVTT）所忽略的空间推理、时间动态、组合理解和长上下文检索等方面的关键差距。

UVRB is a comprehensive benchmark suite designed to evaluate the generalization ability of video embedding models, containing 16 sub-datasets supporting 3 query types (text-to-video, composed query, visual query), and assessed across 6 capability dimensions. It not only focuses on accuracy but also on why the model succeeds or fails. UVRB reveals critical gaps in spatial reasoning, temporal dynamics, compositional understanding, and long-context retrieval that traditional benchmarks like MSRVTT completely miss.