Update README.md

THVL-Bench: A Multimodal Benchmark for Temporal Localization of Harmful Content in Long Videos

## Introduction
Video-level safety assessment is often insufficient for real-world moderation because harmful content is typically sparse and temporally localized. We introduce **Temporal Harmful Video Localization (THVL)**, a task requiring models to identify harmful segments with precise temporal boundaries and multimodal rationales.

## Dataset Summary
**THVL-Bench** is a comprehensive benchmark designed to evaluate the capability of Multimodal Large Language Models (MLLMs) in detecting and localizing harmful content within long-form videos.
- **Total Videos**: 450 real-world long videos sourced from YouTube and Bilibili.
- **Annotations**: 1,099 manually annotated segments with precise start/end timestamps.
- **Categories**: 11 distinct harm categories reflecting diverse online safety challenges.
- **Key Feature**: Each segment includes multimodal rationales (visual/audio) explaining the harmfulness.

## Supported Tasks
1. **Temporal Localization**: Identifying the precise interval where harmful content occurs.
2. **Harm Category Classification**: Labeling the specific type of harm (e.g., violence, harassment).
3. **Multimodal Attribution**: Explaining whether the harm originates from visual cues, audio, or both.
4. **Rationale Generation**: Generating natural language explanations for the moderation decision.

## Evaluation
A systematic evaluation of state-of-the-art MLLMs on THVL-Bench reveals a significant "detection-localization gap". While models can often recognize that harm exists, they struggle with precise temporal grounding. We also provide **THVL-Expert**, a multi-expert framework as a baseline to enhance temporal reasoning.

README.md

@@ -1,3 +1,18 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+task_categories:
+- video-classification
+language:
+- en
+tags:
+- temporal-localization
+- video-understanding
+- multimodal
+- multimodal-safety
+- content-moderation
+- long-video
+- benchmark
+pretty_name: THVL-Bench
+size_categories:
+- n<1K
+---