librakevin/lewm-pusht

LeWM PushT (Lance format)

The PushT expert dataset used in LeWorldModel, converted from HDF5 to Lance format.

The original data is quentinll/lewm-pusht (pusht_expert_train.h5.zst). This repository has the same episodes and steps, stored as a Lance dataset. Lance gives fast random row access, which works well with many parallel DataLoader workers.

Contents

pusht_expert_train.lance/   # Lance dataset (directory), ~14 GB

Episodes	18,685
Steps (rows)	2,336,736
Episode length	49 to 246 steps (mean 125)
Image size	224 × 224 RGB

Schema

One row per step:

Column	Type	Description
episode_idx	int32	Episode index (0 to 18684)
step_idx	int32	Step index within the episode
pixels	binary	JPEG-encoded 224×224 RGB frame
action	fixed_size_list<float>[2]	2-D action
proprio	fixed_size_list<float>[4]	Proprioceptive state (4-D)
state	fixed_size_list<float>[7]	Full environment state (7-D)

Conversion

The data was converted with stable-worldmodel 0.1.1 (pylance 12.0.0):

swm convert pusht_expert_train pusht_expert_train.lance -f lance

• episode_idx and step_idx are rebuilt by the Lance writer in the same episode order as the source.
• action, proprio and state are copied exactly.
• pixels are re-encoded as JPEG at quality 95 (about 5.8 KB per frame). The HDF5 source stores them losslessly. This is why the dataset is about 14 GB instead of about 44 GB uncompressed.

JPEG is lossy, but the difference is very small. On 200 random frames, the mean absolute pixel difference from the HDF5 originals was 0.11 out of 255 (PSNR ≈ 53 dB), and episode_idx, step_idx and action matched exactly. If you need bit-exact pixels, use the original HDF5 dataset.

Usage

Load with stable-worldmodel by repo id:

import stable_worldmodel as swm

# Uses $STABLEWM_HOME/datasets/librakevin--lewm-pusht/, downloading it on first use
ds = swm.data.load_dataset("librakevin/lewm-pusht")
print(len(ds.lengths), int(ds.lengths.sum()))  # 18685 episodes, 2336736 steps

Or download it yourself into the same folder, which resumes if interrupted:

hf download librakevin/lewm-pusht --repo-type dataset --local-dir $STABLEWM_HOME/datasets/librakevin--lewm-pusht

Or read directly with Lance:

import io
import lance
from PIL import Image

ds = lance.dataset("pusht_expert_train.lance")
row = ds.take([0]).to_pylist()[0]
frame = Image.open(io.BytesIO(row["pixels"]))  # 224x224 RGB

To train LeWM on it, set dataset.name: librakevin/lewm-pusht in config/train/data/pusht.yaml and point LOCAL_DATASET_DIR at the directory that contains datasets/.

License and citation

MIT, following the source dataset. If you use this data, please cite LeWorldModel (paper) and credit the original dataset, quentinll/lewm-pusht.