What it is
Synthetic video-continuation corpus for training and testing object permanence in world models across hand-designed cognitive tasks. 1.5M Blender-rendered video-continuation samples across 150 hand-designed tasks in 6 cognitive categories, with paired input/target clips, text prompts, trajectories, and metadata.
Key results
- 1.5 million samples across 150 tasks
- 1280x720 videos at 24 fps, 60 input + 60 target frames
- 5 files per sample; CC BY-NC 4.0; 214 GB total
Useful for training and benchmarking video models on object permanence and physical reasoning.