THVL-Bench: A Multimodal Benchmark for Temporal Localization of Harmful Content in Long Videos.
Description: THVL-Bench is a large-scale multimodal benchmark designed for the task of Temporal Harmful Video Localization (THVL). While traditional video safety tools focus on video-level classification, THVL-Bench requires models to identify the exact start and end times of harmful segments within long-form videos and provide multimodal rationales for their decisions . The dataset contains 450 real-world long videos collected from platforms like YouTube and Bilibili, featuring 1,099 manually annotated segments across 11 distinct harm categories .
Dataset Structure: The dataset is provided in a structured format (CSV/Parquet) with the following key columns:
videoID: The unique identifier for the video (e.g., Bilibili BV IDs or YouTube IDs).
segment-level annotation: A list of binary vectors representing the presence of 11 harm categories for each segment.
segment-level timestamp: The precise temporal boundaries [start_time, end_time] for each identified harmful segment.
contributing modalities: A triplet vector [Visual, Audio, Text] indicating which modality contributes to the harmfulness of the segment.
Harm Categories: The benchmark covers 11 fine-grained categories of harmful content. One or more labels from: : ["InformationHarm" , "VerbalAbuse" , "Hate" , "Bias" , AddictionHarm" , "SexualHarm" , "PhysicalHarm" , "Violence" , "Blood/Gore" , "CriminalActivity", "Danger"] (e.g., [0,0,0,1,0,0,0,0,0,0,0]Segment labels correspond to: "Bias")