Planning baseline checkpoint archive
This repository archives checkpoints and measured planning results for TwoRoom, Cube and PushT. Reacher is excluded and LeJEPA is canceled. Original checkpoints retain their upstream licenses; see licenses/.
| Directory | Contents | Status |
|---|---|---|
| fast-lewm/tworoom, cube, pusht | Exact original naiverer/fast-leworldmodel checkpoints |
All five-seed planning evaluations completed |
| dinowm/pusht | Original DINO-WM OSF PushT checkpoint and frozen DINOv2-S/14 backbone | Published model trained on pusht_noise; evaluation incomplete |
| pldm/cube | Resume checkpoint, configuration and provenance | Intermediate checkpoint; read provenance.json for completed epoch count, not yet the final 10-epoch model |
| pldm/tworoom | Configuration and results only | The trained weights were deleted after evaluation under the earlier disk-saving instruction and are unavailable |
Newly completed local checkpoints are archived under trained/<method>/<task>/model.pt, with provenance alongside each file. Each local checkpoint uses training seed 42. Five evaluation seeds do not mean five independently trained models.
Results and comparison protocol
results/planning_success.xlsx contains the detailed per-seed results and the seed summary. results/planning_summary.csv contains columns for seeds 42, 43, 44, 45, 46 and mean/sample standard deviation (ddof=1). Success percentages are reported only for 200 completed episodes per seed and goal offset. Goal offsets are 25, 50, 75 and 100. Mean and std require all five completed seeds; missing values are null.
The planning budget is 50 environment steps, with a 25-step planning window and CEM using 300 candidates, 30 iterations and 30 elites. Models share the persisted task/offset/seed evaluation pairs. Newly trained PLDM and DINO-WM checkpoints use a fixed 10-epoch training budget, while published Fast-LeWM and DINO-WM checkpoints have their original training histories. This is not an equal-training-budget comparison or an exact paper reproduction. Evaluation pairs come from the supplied dataset and are not a separate held-out test split.
Loading
Original Fast-LeWM object checkpoints depend on the upstream Fast-LeWorldModel Python source. Original DINO-WM objects depend on the original DINO-WM source. Locally trained objects additionally depend on code/train_models.py and stable-worldmodel 0.1.1. The Cube resume checkpoint holds state dictionaries, optimizer/RNG state and completed-epoch metadata; it is not a directly callable model object. See code/train_baseline.py for restoration. Loading pickled Python model objects requires their matching source and a trusted checkpoint source.
Source repositories:
- https://github.com/lucas-maes/le-wm
- https://huggingface.co/naiverer/fast-leworldmodel
- https://github.com/gaoyuezhou/dino_wm
No source datasets are uploaded. No paper-reported number is substituted for a measured score. For complete protocol details and known implementation differences, see results/PROTOCOL.md.