THVL/THVL-Bench

Dataset Card for Dataset Name

THVL-Bench is a multimodal benchmark for Temporal Harmful Video Localization (THVL), designed to evaluate state-of-the-art Multimodal Large Language Models (MLLMs) on their ability to precisely localize harmful content in long-form videos, with fine-grained category labels and modality contribution annotations. The dataset contains 450 real-world long videos with 1,099 manually annotated harmful segments across 11 harm categories.

This dataset card aims to be a base template for new datasets. It has been generated using this raw template.

Dataset Details

Dataset Description

THVL-Bench is the first benchmark dedicated to the Temporal Harmful Video Localization (THVL) task, which requires models to jointly perform harmful event detection, precise temporal boundary localization, fine-grained harm category classification, and contributing modality attribution for long untrimmed videos. Unlike conventional video-level harmful content classification datasets, THVL-Bench focuses on the realistic moderation scenario where harmful content is temporally sparse and embedded in extended contextual content. The dataset consists of 450 real-world long videos (all ≥3 minutes in duration) collected from YouTube and Bilibili, with 1,099 manually annotated harmful segments. Each segment is annotated with:

1. Precise start/end timestamps
2. One or more labels from 11 fine-grained harmful categories
3. Contributing modality labels (visual, textual, auditory)
4. Auxiliary annotation rationales

• Curated by: [authors of the THVL-Bench paper]
• Funded by [optional]: [THVL-Bench Authors]
• Shared by [optional]: [THVL-Bench Authors]
• Language(s) (NLP): [Engish]
• License: [CC BY-NC 4.0]

Dataset Sources [optional]

• Repository: [More Information Needed]
• Paper [optional]: [THVL-Bench: A Multimodal Benchmark for Temporal Localization of Harmful Content in Long Videos]
• Demo [optional]: [N/A]

Uses

Direct Use

• Benchmark Evaluation: Systematic evaluation of state-of-the-art MLLMs on temporal harmful content localization, detection, and modality attribution tasks
• Model Development: Training and fine-tuning multimodal models for video safety understanding, harmful content moderation, and temporal reasoning
• Research Analysis: Studying the detection-localization gap of MLLMs, cross-category performance variation, and multimodal dependency patterns in harmful content understanding
• Safety System Validation: Testing the robustness of video moderation systems on realistic long-form video scenarios

Out-of-Scope Use

• Malicious Use: Using the dataset to generate, distribute, or promote harmful content
• Unauthorized Redistribution: Redistributing the raw video files (not included in this dataset release) in violation of original platform terms of service
• Commercial Use: Using the dataset for commercial purposes without explicit authorization from the authors
• Unrelated Tasks: Using the dataset for tasks outside of video safety, harmful content understanding, and temporal localization research
• Biased Model Training: Training models that perpetuate harmful stereotypes or biases using the dataset

Dataset Structure

The dataset is released as a single CSV file with 450 rows (one row per video) and 4 core columns, with segment-level annotations stored in structured list format:

Column Name	Description
videoID	Unique identifier for each video, corresponding to the original platform's video ID
segment-level annotation	One-hot encoded labels for 11 harmful categories, with one list per annotated segment in the video
segment-level timestamp	Start and end timestamps (in mm:ss format) for each annotated harmful segment, with one pair per segment
contributing modalities	One-hot encoded labels for 3 contributing modalities (visual, textual, auditory), with one list per annotated segment

Key Dataset Statistics

• Total videos: 450
• Total annotated harmful segments: 1,099
• Video duration range: 3–40 minutes
• Average harmful segment duration: 68.31 seconds
• Harmful categories: 11 fine-grained classes (Information Harm, Verbal Abuse, Hate, Bias, Addiction Harm, Sexual Harm, Physical Harm, Violence, Blood/Gore, Criminal Activity, Danger)
• Contributing modalities: 3 classes (Visual, Textual, Auditory)

Dataset Creation

Curation Rationale

Existing multimodal safety benchmarks primarily focus on video-level harmfulness recognition or safe response generation, rather than precise temporal localization of harmful content in long videos. This creates a critical gap: it remains unclear whether current MLLMs can reliably support temporally grounded harmful video moderation in real-world scenarios, where harmful content is often sparse and embedded in extended contextual content.

THVL-Bench was created to address this gap by:

• Defining the THVL task, which requires joint detection, temporal localization, category classification, and modality attribution
• Providing a high-quality, manually annotated benchmark of real-world long videos
• Enabling systematic evaluation of MLLMs' capabilities and limitations in temporally precise harmful content understanding

Source Data

All videos in THVL-Bench are collected from publicly accessible platforms, including YouTube and Bilibili, for research purposes only. The dataset focuses on long-form untrimmed videos (all ≥3 minutes in duration) to reflect realistic moderation scenarios, where harmful events are often temporally sparse relative to the full video duration.

Data Collection and Processing

• Taxonomy Definition: Adopted and adapted existing online harm taxonomies from prior safety and harmful content research to define 11 fine-grained harmful categories
• Keyword Search: Manually constructed category-specific keyword sets and search phrases associated with harmful behaviors, dangerous activities, hate expressions, offensive interactions, criminal events, and sensitive social content
• Candidate Retrieval: Used the constructed keywords to retrieve candidate videos from the search engines of YouTube and Bilibili
• Manual Filtering: Candidate videos were manually reviewed to remove duplicated, inaccessible, low-quality, or irrelevant samples
• Duration Filtering: Only videos exceeding 3 minutes in duration were retained to focus on long-form video scenarios

Source Data Producers

The source videos are publicly available content created by users of YouTube and Bilibili. The dataset release only includes annotations and metadata, not the raw video files, in compliance with the original platforms' terms of service.

Annotations

Annotation Process

• Annotator Training: All annotators underwent a calibration phase using 20 pilot videos before formal annotation, with clear guidelines on temporal boundary marking, category labeling, and modality attribution.
• Dual Annotation: Each video was independently annotated by two individuals, who identified harmful events, marked precise start/end timestamps, assigned category labels, identified contributing modalities, and wrote brief annotation rationales.
• Disagreement Resolution: Disagreements regarding temporal boundaries, categories, or modalities were first addressed through discussion; unresolved cases were sent for a second round of review, with final labels determined by majority voting if consensus was not reached.
• Quality Control: Inter-annotator agreement was measured on a randomly sampled subset, achieving a Cohen’s κ=0.74 for segment localization, with 92.3% agreement on category and modality labels.

Who are the annotators?

THVL-Bench was annotated by four experts with extensive experience in multimodal video analysis and content understanding. All annotators were informed of the potential risks prior to their participation and had unrestricted access to institutional psychological counseling resources throughout the annotation process.

Personal and Sensitive Information

This dataset release does not contain any raw video files, personal identifiable information, or private sensitive data. Only anonymized video IDs, segment-level annotations, timestamps, and modality labels are included. All annotations are aggregated at the segment level with no reference to specific individuals in the videos. The dataset is intended for research purposes only, and all users are expected to comply with the original platforms' terms of service and use the data responsibly.

Bias, Risks, and Limitations

Limitations

• Scale Constraints: The current benchmark contains 450 videos with 1,099 annotated segments, which is relatively limited compared to large-scale general video datasets.
• Category Distribution Imbalance: Some harmful categories (e.g., Hate, Addiction Harm, Physical Harm) have fewer annotated samples than more prevalent categories (e.g., Danger, Violence, Criminal Activity), which may affect model evaluation on underrepresented classes.
• Limited Context Modeling: Annotations primarily focus on temporal grounding and modality attribution, while higher-level contextual factors such as intent, social context, and cultural nuance are not fully modeled.
• Platform Bias: Videos are collected from YouTube and Bilibili, which may not fully represent harmful content scenarios on other video platforms or in other regions.

Risks

• Sensitive Content Exposure: The dataset describes and annotates harmful content including violence, hate speech, criminal activities, and other unsafe behaviors, which may be disturbing to some users.
• Potential Misuse: The dataset could be misused to develop or optimize systems that generate or distribute harmful content.
• Bias Amplification: Models trained on the dataset may amplify existing biases in the annotation data or source content if not properly validated.

Recommendations

Users should be made aware of the risks, biases and limitations of the dataset. Key recommendations include:

• Responsible Use: Use the dataset exclusively for research purposes related to video safety, harmful content understanding, and multimodal AI safety.
• Bias Mitigation: When training models on the dataset, implement bias mitigation techniques to address category distribution imbalance and avoid amplifying harmful stereotypes.
• Ethical Review: Conduct ethical review of any systems or models developed using the dataset before deployment in real-world scenarios.
• Compliance: Comply with the terms of service of the original video platforms and the CC BY-NC 4.0 license of the dataset.
• Mental Wellbeing: Take appropriate precautions when working with the dataset, as it describes sensitive and potentially disturbing harmful content.

Users should be made aware of the risks, biases and limitations of the dataset.

Citation [optional]

BibTeX:

[More Information Needed]

APA:

[More Information Needed]

Glossary [optional]

[More Information Needed]

More Information [optional]

[More Information Needed]

Dataset Card Authors [optional]

[More Information Needed]

Dataset Card Contact

[More Information Needed]