EchoRaven/safewatch-genai

SafeWatch-TV2V Dataset

Overview

This dataset combines SafeWatch-Bench original videos and TV2V generated videos for video safety detection and robustness testing.

Dataset Structure

Three Main Components:

1. TV2V-output: Generated videos with prompts

   • Videos: 1,040 files
   • Metadata: 1,213 entries in TV2V-output.jsonl
   • Content: Generated videos with their corresponding prompts

2. TV2V-input: Original videos used for generation

   • Videos: 410 files (deduplicated)
   • Metadata: 1,213 entries in TV2V-input.jsonl
   • Content: Original SafeWatch-Bench videos used as input for TV2V generation

3. Safewatch-Bench-Genai: SafeWatch-Bench original genai videos

   • Videos: 590 files
   • Metadata: 590 entries in Safewatch-Bench-Genai.jsonl
   • Content: Original SafeWatch-Bench genai videos

Data Format

Each JSONL entry contains:

{
  "category": "C1",
  "subcategory": "evident", 
  "video_path": "TV2V-output/output_chunk_0_1_20250611_222842.mp4",
  "video_content": "The frame shows a woman lying sideways in bed..."
}

Categories

• C1: Sexual Content
• C2: Abuse Content
• C3: Violence Content
• C4: Misinformation Content
• C5: Criminal Content
• C6: Harmful Content

Video Content Sources

• TV2V-output video_content: Direct from TV2V output.jsonl prompt field
• TV2V-input video_content: From SafeWatch-Bench dataset video_content field
• Safewatch-Bench-Genai video_content: From SafeWatch-Bench dataset video_content field

Usage

import json

# Load dataset
with open('TV2V-output.jsonl', 'r') as f:
    for line in f:
        data = json.loads(line)
        print(f"Video: {data['video_path']}")
        print(f"Content: {data['video_content']}")

Statistics

• Total videos: 2,040 files
• Total metadata entries: 3,016 entries
• TV2V generated videos: 1,213 entries (1,040 unique files)
• SafeWatch-Bench videos: 590 entries (590 files)

License

MIT License