README.md

---
license: mit
task_categories:
- reinforcement-learning
- robotics
tags:
- world-models
- le-wm
- imitation-learning
- arxiv:2603.19312
pretty_name: LeWM Original Benchmark Datasets (PushT / Cube / TwoRooms / Reacher)
---

# LeWM — Original Benchmark Datasets (4-in-one)

The four original benchmark datasets from **LeWorldModel (LeWM)**
([project page](https://le-wm.github.io/), arXiv:2603.19312), collected in a single repo
for convenience. These are the **pure-original published archives**, byte-identical
(SHA-256 verified) to the upstream `quentinll/lewm-*` sources.

| Task | File | Size | Format |
|---|---|---|---|
| PushT | `pusht_expert_train.h5.zst` | 13.1 GB | zstd-compressed swm-HDF5 |
| Cube (single) | `cube_single_expert.tar.zst` | 46.2 GB | zstd `.tar` archive |
| TwoRooms | `tworoom.tar.zst` | 3.4 GB | zstd `.tar` archive |
| Reacher | `reacher.tar.zst` | 23.8 GB | zstd `.tar` archive |

## Provenance

Mirrored from the authoritative upstream repos, each file content-hash (SHA-256) verified
against its source:

- `quentinll/lewm-pusht` → `pusht_expert_train.h5.zst`
- `quentinll/lewm-cube` → `cube_single_expert.tar.zst`
- `quentinll/lewm-tworooms` → `tworoom.tar.zst`
- `quentinll/lewm-reacher` → `reacher.tar.zst`

## Decompress

```bash
# PushT — a single .h5 inside the zstd wrapper
zstd -d pusht_expert_train.h5.zst

# the three tar.zst archives
tar --use-compress-program=unzstd -xf cube_single_expert.tar.zst
tar --use-compress-program=unzstd -xf tworoom.tar.zst
tar --use-compress-program=unzstd -xf reacher.tar.zst
```

PushT is directly loadable by [stable-worldmodel](https://github.com/galilai-group/stable-worldmodel)
as swm-HDF5; the other three are the raw collected archives.

## License

MIT, inherited from the upstream LeWM release.