Dataset · TIGER-Lab
Generated Video Quality Ratings Across 5 Criteria
TIGER-Lab/VideoFeedback
37.6 thousand text-video pairs from 11 generators with human annotations for training evaluation models.
What's inside
The TIGER-Lab dataset includes 37,600 text prompt-video pairs generated by 11 popular video models, along with real clips. Raters evaluated each video on a scale from 1 to 4 across five parameters: visual quality, temporal consistency, motion, text alignment, and factual accuracy. The dataset was used to train the VideoScore evaluation model.
What you can do with it
- Training reward models and automated video generation quality metrics (such as VideoScore)
- RLHF and alignment of text-to-video generative models
- Analyzing common artifacts and weaknesses of modern video generators
- Tasks
- Video classification
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 13.1K
- Likes
- 35
Models trained on it
How to load it
from datasets import load_dataset
ds = load_dataset("TIGER-Lab/VideoFeedback")