Autobrik/teaching-cartpole-10k-v1

TeachingCartpole 10k (v1)

Offscreen-rendered single-pole CartPole episodes (RGB frames, physical state, and the applied horizontal force) in the stable-worldmodel HDF5 layout used to train LeWorldModel. Collected for COMP 765 (McGill) Assignment 1, Q2(e). Models trained on it: Autobrik/lewm-teaching-cartpole-fs5, Autobrik/lewm-teaching-cartpole-fs1.

Contents

10,000 episodes, 1,507,974 transitions, split by episode (seed 43):

File	Episodes	Noisy LQR	Random	Transitions
train.h5	8,000	3,989	4,011	1,207,292
validation.h5	1,000	490	510	147,680
test.h5	1,000	521	479	153,002

End reasons: 5,679 time limit (200 transitions), 4,321 track/camera bound (|x| ≥ 2.5 m). manifest.json records per-episode seeds, initial states, policy, split, and end reason, plus physics constants and source hashes.

HDF5 column	Shape / content
pixels	(N, 224, 224, 3) uint8 RGB, fixed camera, gzip
action	(N, 1) float32 horizontal force in N, ±10.0
observation	(N, 4) float64 ['x', 'x_dot', 'theta_dot', 'theta'] (m, m/s, rad/s, rad)
timestamp	(N,) simulated seconds from reset
episode_id, terminated, truncated	per row
ep_len, ep_offset	rows and start row of each episode

• Alignment: action[t] advances observation[t] to observation[t+1]; last is NaN. An episode with T transitions has T+1 rows. Never treat the final NaN as a zero-force action.
• θ = 0 is hanging down, θ = π upright; angles are unwrapped.
• Control at 20 Hz (dt 0.05 s = 10 physics substeps of 0.005 s). Physics: g 9.82, cart M 0.5 kg, pole m 0.5 kg, l 0.5 m, cart viscous friction b 1.0 N·s/m (force − b·ẋ). Frames are stored at every control step; subsample at load time with the loader's frameskip.
• Policies: noisy local LQR balancing from near upright (Gaussian force noise σ 0.5 N) and uniform random forces held 2 steps from mixed initial angles. Little swing-up behavior is present.

Loading

import hdf5plugin, stable_worldmodel as swm
from huggingface_hub import snapshot_download

d = snapshot_download("Autobrik/teaching-cartpole-10k-v1", repo_type="dataset")
ds = swm.data.load_dataset(f"{d}/train.h5", frameskip=5, num_steps=4,
                           keys_to_load=["pixels", "action", "observation"])

Pass an absolute path (a relative one is resolved as a Hub repo id). action and observation are stored with one-row chunks: reading a whole column at once (e.g. the loader's keys_to_cache) costs ~8 GB of HDF5 bookkeeping for train.h5. Read them in slices of ~50k rows instead (identical arrays, <0.5 GB), as the fork's eval_lewm.py does.

Validation

Every episode's structure and boundaries were checked; a physical replay of 10 episodes per split reproduces the stored states; the stable-worldmodel native reader was checked on the first/last windows of every split (validation/). Collector: fork BrikHMP18/le-wm, branch feat/teaching-cartpole-data, commit c628900 (experiments/teaching-cartpole).