---
license: mit
tags:
- stable-worldmodel
- pusht
- smoke-test
---

# PushT Sample

Tiny Stable World Model / LeWM-compatible PushT HDF5 sample for CUDA and training-loop smoke tests.

This is not an evaluation dataset. It contains two short real PushT episodes with 40 frames each, enough for one default-sized training batch with the RLC 2026 LeWM config (`frameskip=5`, `history_size=3`, `num_preds=1`).

Use with the workshop repo via:

```bash
python train_eggroll.py data=pusht_sample \
  eggroll.device=cuda \
  eggroll.steps=5 \
  eggroll.population=2 \
  loader.batch_size=8 \
  eval_loader.batch_size=1 \
  eggroll.planning.enabled=false
```

For planning-smoke tests, also set:

```bash
eggroll.planning.dataset_name=fracapuano/pusht-sample
```