ngqtrung/raw_set_3372

Raw Video Dataset (3372 Videos)

This dataset contains 3372 videos with both single-modality and cross-modality question-answering annotations.

Dataset Structure

train/
├── metadata.json                 # Combined metadata with all annotations
├── videos_part_001.tar          # Video files (part 1)
├── videos_part_002.tar          # Video files (part 2)
└── ...                          # Additional parts (~5GB each)

Metadata Format

The metadata.json file contains:

{
  "video_id": {
    "video_path": "video_id.mp4",
    "single_modality": {
      "vision_only": { "question": "...", "choices": {}, "correct_answer": "..." },
      "vision_only_misleading": { ... },
      "audio_only": { ... },
      "audio_only_misleading": { ... }
    },
    "cross_modality": {
      "task0": { "variant_type": "default", "question": "...", ... },
      "task1": { "variant_type": "audio_misleading", ... },
      "task2": { "variant_type": "visual_misleading", ... }
    }
  }
}

Question Types

Single Modality

• vision_only: Questions about visual content only
• vision_only_misleading: Vision questions with misleading visual information
• audio_only: Questions about audio content only
• audio_only_misleading: Audio questions with misleading audio information

Cross Modality

• default: Questions requiring both audio and visual understanding
• audio_misleading: Cross-modal questions with misleading audio
• visual_misleading: Cross-modal questions with misleading visuals

Options

• Questions include options A, B, C, D
• Option E: "Vision details are wrong" (for vision questions) or "Audio details are wrong" (for audio questions)
• Option F: "Audio details are wrong" (only for cross-modality questions)

Usage

Extract Videos

# Extract all tar files
for tar_file in train/videos_part_*.tar; do
    tar -xf "$tar_file" -C videos/
done

Load Metadata

import json

with open('train/metadata.json', 'r') as f:
    metadata = json.load(f)

# Access data for a video
video_id = "example_video_id"
video_data = metadata[video_id]
print(video_data['single_modality']['vision_only']['question'])

Statistics

• Total Videos: 3372
• Total Tar Files: 18
• Single Modality Questions: 13488
• Cross Modality Questions: 10116

Citation

If you use this dataset, please cite appropriately.

License

Please check the original video sources for licensing information.