Skip to content

Dataset · osazuwa

WHAM 3D simulation scenario video rollouts

osazuwa/wham3d-cts-trace-rollouts-20260923

Cached video rollouts of scene continuations for diagnosing and analyzing the behavior of the WHAM 3D model.

What's inside

The dataset contains between 1 and 10 thousand video recordings, which are cached rollouts for given initial fragments at step 500 of model training. For each starting prefix, two scenarios are recorded: normal observation without intervention and diagnostic observation with the do(X=1) intervention. The dataset is growing and stratified-balanced.

What you can do with it

  • Diagnosing causal relationships and generative video model behavior under interventions
  • Analyzing and debugging prediction trajectories in 3D environments
Size
1K–10K rows
Downloads this month
20K
Likes
0

How to load it

from datasets import load_dataset

ds = load_dataset("osazuwa/wham3d-cts-trace-rollouts-20260923")