full-modality-bench / README.md
ngqtrung's picture
Update README.md
b3563c2 verified
---
license: mit
task_categories:
- video-classification
- question-answering
- visual-question-answering
language:
- en
size_categories:
- 1K<n<10K
---
# Multimodal Video QA Dataset
This dataset contains challenging video question-answering tasks that require understanding both visual and audio information across entire video timelines.
## Dataset Statistics
- **Total Videos**: 4,140
- **Total Size**: 118.08 GB
## Dataset Structure
The dataset is split into multiple parts (each ≤2GB):
- **Part 1**: 75 videos (1.97 GB) - `videos_part001.zip`
- **Part 2**: 49 videos (1.97 GB) - `videos_part002.zip`
- **Part 3**: 55 videos (1.91 GB) - `videos_part003.zip`
- **Part 4**: 82 videos (1.99 GB) - `videos_part004.zip`
- **Part 5**: 81 videos (1.98 GB) - `videos_part005.zip`
- **Part 6**: 66 videos (1.98 GB) - `videos_part006.zip`
- **Part 7**: 56 videos (1.78 GB) - `videos_part007.zip`
- **Part 8**: 66 videos (1.99 GB) - `videos_part008.zip`
- **Part 9**: 62 videos (1.97 GB) - `videos_part009.zip`
- **Part 10**: 70 videos (2.00 GB) - `videos_part010.zip`
- **Part 11**: 68 videos (1.99 GB) - `videos_part011.zip`
- **Part 12**: 83 videos (1.97 GB) - `videos_part012.zip`
- **Part 13**: 59 videos (1.89 GB) - `videos_part013.zip`
- **Part 14**: 58 videos (1.96 GB) - `videos_part014.zip`
- **Part 15**: 79 videos (1.97 GB) - `videos_part015.zip`
- **Part 16**: 60 videos (1.97 GB) - `videos_part016.zip`
- **Part 17**: 70 videos (1.97 GB) - `videos_part017.zip`
- **Part 18**: 80 videos (2.00 GB) - `videos_part018.zip`
- **Part 19**: 97 videos (1.99 GB) - `videos_part019.zip`
- **Part 20**: 78 videos (1.95 GB) - `videos_part020.zip`
- **Part 21**: 81 videos (2.00 GB) - `videos_part021.zip`
- **Part 22**: 68 videos (1.99 GB) - `videos_part022.zip`
- **Part 23**: 61 videos (1.99 GB) - `videos_part023.zip`
- **Part 24**: 66 videos (1.92 GB) - `videos_part024.zip`
- **Part 25**: 61 videos (1.90 GB) - `videos_part025.zip`
- **Part 26**: 66 videos (1.97 GB) - `videos_part026.zip`
- **Part 27**: 83 videos (1.97 GB) - `videos_part027.zip`
- **Part 28**: 55 videos (1.99 GB) - `videos_part028.zip`
- **Part 29**: 84 videos (1.95 GB) - `videos_part029.zip`
- **Part 30**: 60 videos (1.94 GB) - `videos_part030.zip`
- **Part 31**: 63 videos (1.85 GB) - `videos_part031.zip`
- **Part 32**: 59 videos (1.95 GB) - `videos_part032.zip`
- **Part 33**: 58 videos (1.96 GB) - `videos_part033.zip`
- **Part 34**: 57 videos (1.89 GB) - `videos_part034.zip`
- **Part 35**: 64 videos (1.97 GB) - `videos_part035.zip`
- **Part 36**: 72 videos (2.00 GB) - `videos_part036.zip`
- **Part 37**: 74 videos (1.99 GB) - `videos_part037.zip`
- **Part 38**: 64 videos (1.97 GB) - `videos_part038.zip`
- **Part 39**: 77 videos (1.98 GB) - `videos_part039.zip`
- **Part 40**: 66 videos (1.80 GB) - `videos_part040.zip`
- **Part 41**: 58 videos (1.93 GB) - `videos_part041.zip`
- **Part 42**: 47 videos (1.97 GB) - `videos_part042.zip`
- **Part 43**: 69 videos (1.95 GB) - `videos_part043.zip`
- **Part 44**: 69 videos (1.97 GB) - `videos_part044.zip`
- **Part 45**: 73 videos (1.93 GB) - `videos_part045.zip`
- **Part 46**: 72 videos (1.98 GB) - `videos_part046.zip`
- **Part 47**: 73 videos (2.00 GB) - `videos_part047.zip`
- **Part 48**: 55 videos (2.00 GB) - `videos_part048.zip`
- **Part 49**: 68 videos (1.99 GB) - `videos_part049.zip`
- **Part 50**: 79 videos (1.99 GB) - `videos_part050.zip`
- **Part 51**: 52 videos (1.80 GB) - `videos_part051.zip`
- **Part 52**: 86 videos (1.98 GB) - `videos_part052.zip`
- **Part 53**: 88 videos (1.99 GB) - `videos_part053.zip`
- **Part 54**: 74 videos (1.97 GB) - `videos_part054.zip`
- **Part 55**: 62 videos (1.98 GB) - `videos_part055.zip`
- **Part 56**: 81 videos (1.98 GB) - `videos_part056.zip`
- **Part 57**: 75 videos (1.93 GB) - `videos_part057.zip`
- **Part 58**: 70 videos (2.00 GB) - `videos_part058.zip`
- **Part 59**: 69 videos (1.95 GB) - `videos_part059.zip`
- **Part 60**: 68 videos (1.98 GB) - `videos_part060.zip`
- **Part 61**: 19 videos (0.64 GB) - `videos_part061.zip`
**Metadata**: All video metadata is in a single file `metadata.json`
## Question Types
The dataset includes 8 question types testing global video understanding:
1. **Temporal**: Questions about timing and sequence across the video
2. **Causal**: Cause-effect relationships spanning multiple segments
3. **Plot**: Overall narrative arc from beginning to end
4. **Cross-modality**: Patterns emerging from visual + audio combination
5. **Emotional**: Emotional journeys across the timeline
6. **Time Order**: Sequence of major events
7. **Existence**: Recurring patterns throughout the video
8. **Scene Description**: Scene progression across segments
## Question Variants
Each video has 3 question variants:
1. **Default**: All information is correct (4 choices: A-D)
2. **Audio Misleading**: Subtle errors in sound/speech information (5 choices: A-E, correct answer is E "None of the above")
3. **Visual Misleading**: Subtle errors in visual/action information (5 choices: A-E, correct answer is E "None of the above")
## Metadata Format
The single `metadata.json` file contains all video metadata:
```json
{
"video_id_1": {
"video_id": "video_id_1",
"video_duration": "120s",
"duration": "1-2min",
"num_segments": 12,
"question_type": "temporal",
"task0": {
"variant_type": "default",
"question": "...",
"answer": "B",
"candidates": ["A. ...", "B. ...", "C. ...", "D. ..."],
"reasoning": "..."
},
"task1": {
"variant_type": "audio_misleading",
"question": "...",
"answer": "E",
"candidates": ["A. ...", "B. ...", "C. ...", "D. ...", "E. None of the above"],
"reasoning": "..."
},
"task2": {
"variant_type": "visual_misleading",
"question": "...",
"answer": "E",
"candidates": ["A. ...", "B. ...", "C. ...", "D. ...", "E. None of the above"],
"reasoning": "..."
}
},
"video_id_2": {
...
}
}
```
## Usage
```python
import json
import zipfile
# Load metadata for ALL videos
with open('metadata.json', 'r') as f:
metadata = json.load(f)
# Extract all video zip files
for zip_file in ['videos_part001.zip', 'videos_part002.zip', ...]:
with zipfile.ZipFile(zip_file, 'r') as zip_ref:
zip_ref.extractall('videos/')
# Access QA data
for video_id, data in metadata.items():
print(f"Video: {video_id}")
print(f"Question: {data['task0']['question']}")
print(f"Answer: {data['task0']['answer']}")
# Or access specific video
video_data = metadata['specific_video_id']
print(f"Default question: {video_data['task0']['question']}")
print(f"Audio misleading: {video_data['task1']['question']}")
print(f"Visual misleading: {video_data['task2']['question']}")
```
## Citation
If you use this dataset, please cite:
```bibtex
@dataset{multimodal_video_qa,
title={Multimodal Video QA Dataset},
year={2025},
publisher={HuggingFace}
}
```
## License
MIT License