The first comprehensive benchmark for universal video retrievalEvaluate your model across 16 datasets, 3 query types, and 6 capability dimensions — not just accuracy, but why it succeeds or fails.
UVRB is a comprehensive evaluation suite designed to diagnose and quantify a video embedding model’s true generalization ability — beyond narrow text-to-video tasks. It exposes critical gaps in spatial reasoning, temporal dynamics… See the full description on the dataset page:
https://huggingface.co/datasets/Alibaba-NLP/UVRB.
