jiomimi/UniAR-Bench

📂 UniAR-Bench: Benchmarking Evidence-Grounded Anomaly Reasoning Across Modalities in MLLMs UniAR-Bench is a comprehensive cross-modal anomaly reasoning benchmark with extensive tasks and credible evaluation. Built from publicly available datasets, UniAR-Bench comprises over 40,199 question-answer pairs covering coarse-to-fine reasoning across video, image and text modalities organized into four shared challenges: Global and Referring Anomaly Discrimination Fine-Grained Anomaly… See the full description on the dataset page: https://huggingface.co/datasets/jiomimi/UniAR-Bench.