ViSeHate: A Large-Scale Benchmark Dataset for Hate Detection and Temporal Localization in Multimodal Videos

The rapid rise of short-video platforms has accelerated the spread of multimodal hate speech characterized by covert and semantically complex cues. Existing datasets struggle to support real-world content moderation due to limited scale, single-platform bias, and the absence of fine-grained temporal localization annotations. To address these limitations, we introduce ViSeHate, a large-scale and cross-platform benchmark dataset designed for both video-level hate detection and frame-level localization. The dataset comprises ViSeHate-Det (10,000 videos across four platforms and six protected attributes) and ViSeHate-Loc (1,200 videos with precise frame-level boundaries). 

This upload of the ViSeHate-Loc dataset contains 1200 videos with specific time segments and labels.

短视频平台的快速崛起，加速了以隐蔽性、语义复杂性为典型特征的多模态仇恨言论（multimodal hate speech）的传播。现有数据集因规模有限、存在单平台偏差且缺乏细粒度时间定位标注，难以支撑真实场景下的内容审核工作。为解决上述局限，我们推出了ViSeHate，一款兼具大规模与跨平台特性的基准数据集，可同时用于视频级仇恨检测与帧级定位任务。该数据集包含ViSeHate-Det与ViSeHate-Loc两个子集：ViSeHate-Det涵盖覆盖六大受保护属性的四大平台共10000条视频，ViSeHate-Loc则包含带有精确帧级边界的1200条视频。
本次上传的ViSeHate-Loc数据集包含1200条带有特定时间片段与对应标签的视频。