ViolenceDET-Video-Preview

ViolenceDET-Video-Preview is a video understanding safety dataset consisting of 1,796 annotated video clips curated from public sources and benchmark datasets. It is built to detect interpersonal physical violence, assault, and aggressive physical altercations.

The dataset provides raw video files, sequential sampled keyframes (sliced_frames), and structured JSON textual responses (frames_response) containing descriptive captions, step-by-step reasoning, and binary safety and violence indicators.

• Curator: prithivMLmods
• Total Samples: 1,796 rows
• Total Size: ~2.21 GB
• Format: Parquet (video, frames_response, sliced_frames)
• Split: Train (1.3k+ rows)

Dataset Structure & Schema

Feature Fields

Field	Type	Description
video	Video	Source video clip showing physical interaction
sliced_frames	Sequence[Image]	Chronologically ordered keyframes extracted from the video
frames_response	string (JSON)	Structured JSON response detailing caption, classification tags, and rationale

Data Instance Example

{
  "caption": "A video sequence showing two men in a wood-paneled room. One man stands still looking at his phone while another approaches him from behind, initiating an aggressive physical confrontation that escalates into a struggle involving pushing and grappling.",
  "is_violence": true,
  "reason": "The image depicts a clear progression of physical aggression, including a chokehold, pushing against a wall, and general grappling during a fight. This constitutes a 'violent content' violation based on the criteria for physical assault and fighting.",
  "violence": 1,
  "safe": 0
}