370 frozen test cases for video world models, plus the videos each evaluated
model generated for them.
Every case is one first frame + one instruction + one control program.
T4 and T6 additionally carry a scoring checklist. Every first frame passed an
audit (majority of 3 independent gemini-3.1-pro draws) before the case was
frozen; no case is shipped with a known defect.
The three tasks… See the full description on the dataset page: https://huggingface.co/datasets/VWMG/haoyu.
