ModelÜberprüftExtern

Lewm Billiards

> Author: Santosh Jaiswal (@hellojais) > Base architecture: LeWM by Lucas Maes et al. (2025) > Training data: hellojais/billiards-worldmodel > Code: hellojais/le-wm

Quellehuggingface.co/hellojais/lewm-billiards/tree/dc692cbd54e8bc706215a5efd8fccc9d489e1c41

ArtCheckpointBasismodelllucas-maes/le-wmVersionRevision dc692cbVerleger@hellojaisCBewertung

Modellquelle

Art: Checkpoint, Embedding
Dateigröße: 37,0–68,6 MB
Basismodell: lucas-maes/le-wm
Version: vRevision dc692cb
Aktualisiert: 3. Okt. 2026
Verleger: @hellojais
Downloads: 0
Quelle: Hugging Face

Quellenbeschreibung

LeWM Billiards — Trained World Model

Author: Santosh Jaiswal (@hellojais)
Base architecture: LeWM by Lucas Maes et al. (2025)
Training data: hellojais/billiards-worldmodel
Code: hellojais/le-wm

Model variants

File	embed_dim	Input	λ_aux	Best epoch	val/pred_loss	Notes
`lewm_epoch_8_object.ckpt`	192	3-ch	—	8	0.00946	Original full-size transformer
`lewm_small_epoch_8_object.ckpt`	32	3-ch	—	8	0.00280	Transformer baseline
`lewm_mamba_best_object.ckpt`	32	3-ch	—	best	0.00340	Mamba predictor
`lewm_framestacked_best_object.ckpt`	32	9-ch	—	7	0.00594	Frame-stacking; JEPA eviction occurs
`lewm_auxloss_full_best_object.ckpt`	32	9-ch	0.1	7	0.00105	Aux state supervision; eviction fixed; best model

What this model learned

Trained on 4,000 episodes (971,321 frames) of 2D billiards gameplay. The model learned to predict future frame embeddings from current embeddings and actions — encoding billiards physics purely from pixels.

Probe results (linear probe on encoder representations; 192-dim = pre-projector CLS token, 32-dim = post-projector):

Model	Rep dim	pos R2	vel R2
`lewm_small`	32	0.983	0.296
`lewm_mamba`	32	0.983	0.297
`lewm_framestacked`	192	0.446 ⚠️	0.138
`lewm_framestacked`	32	0.599	0.417
`lewm_auxloss_full`	192	0.999 ✅	0.947 ✅
`lewm_auxloss_full`	32	0.982	0.554

Key finding: Frame-stacking causes JEPA representational eviction — the ViT encoder stops encoding ball position (pos R2=0.446 at 192-dim) under optical-flow pressure. Adding a lightweight auxiliary state supervision head (λ=0.1) fully recovers position encoding (pos R2=0.999) and achieves the best prediction loss across all variants.

Planning results

Approach	Same-episode	Novel cross-episode
Pure JEPA embedding CEM	❌ FAIL	❌ FAIL
State-based hybrid CEM	✅ SUCCESS (9 steps)	✅ SUCCESS (13 steps)

Pure JEPA planning failed due to uniform embedding geometry in this visually simple domain. See FINDINGS.md for complete analysis.

Usage

# Load checkpoint
import stable_worldmodel as swm
import torch

device = torch.device("mps")  # or "cuda" or "cpu"

# Load the small model (recommended)
checkpoint = torch.load(
    "lewm_small_epoch_8_object.ckpt", 
    map_location=device
)

Training setup

Hardware: Apple M5 Max (64GB unified memory)
Backend: PyTorch MPS
Training time: ~10–11 hours per run (10 epochs); 4 model variants trained
Framework: PyTorch Lightning + stable-worldmodel

Credits

Original LeWM architecture by: Lucas Maes, Quentin Leroux, Gauthier Gidel, Glen Berseth
Mila / McGill University (2025)
arXiv:2603.19312

Quellen

1 Quelle

Hugging FacePrimärVersion dc692cbd54e8bc706215a5efd8fccc9d489e1c41

huggingface.co/hellojais/lewm-billiards/tree/dc692cbd54e8bc706215a5efd8fccc9d489e1c41

Verifiziert 3. Okt.

Modellartefakte

5 Artefakte

lewm_auxloss_full_epoch_8_object.ckpt

ckpt · 65,9 MB · SHA-256 56e1e58f8516…0574 · Hugging Face

Herunterladen

lewm_epoch_8_object.ckpt

ckpt · 68,6 MB · SHA-256 868f15a76cdb…9e6c · Hugging Face

Herunterladen

lewm_framestacked_best_object.ckpt

ckpt · 65,9 MB · SHA-256 e8006d203938…ebbf · Hugging Face

Herunterladen

lewm_mamba_best_object.ckpt

ckpt · 37,0 MB · SHA-256 f9ef6fc7741b…a25a · Hugging Face

Herunterladen

lewm_small_epoch_8_object.ckpt

ckpt · 65,6 MB · SHA-256 18e6b51c64b4…9765 · Hugging Face

Herunterladen

Quellenauszüge

3 Auszüge

Readme11. Julihuggingface.co/hellojais/lewm-billiards

--- license: mit base_model: lucas-maes/le-wm tags: - world-model - jepa - billiards - pytorch - mps - apple-silicon - frame-stacking - auxiliary-supervision language: - en --- # LeWM Billiards — Trained World Model > **Author:** Santosh Jaiswal ([@hellojais](https://huggingface.co/hellojais)) > **Base architecture:** [LeWM](https://github.com/lucas-maes/le-wm) by Lucas Maes et al. (2025) > **Training data:** [hellojais/billiards-worldmodel](https://huggingface.co/datasets/hellojais/billiards-worldmodel) > **Code:** [hellojais/le-wm](https://github.com/hellojais/le-wm) ## Model variants | File | embed_dim | Input | λ_aux | Best epoch | val/pred_loss | Notes | |---|---|---|---|---|---|---| | `lewm_epoch_8_object.ckpt` | 192 | 3-ch | — | 8 | 0.00946 | Original full-size transformer | | `lewm_small_epoch_8_object.ckpt` | 32 | 3-ch | — | 8 | 0.00280 | Transformer baseline | | `lewm_mamba_best_object.ckpt` | 32 | 3-ch | — | best | 0.00340 | Mamba predictor | | `lewm_framestacked_best_object.ckpt` | 32 | 9-ch | — | 7 | 0.00594 | Frame-stacking; JEPA eviction occurs | | `lewm_auxloss_full_best_object.ckpt` | 32 | 9-ch | 0.1 | 7 | **0.00105** | Aux state supervision; eviction fixed; best model | ## What this model learned Trained on 4,000 episodes (971,321 frames) of 2D billiards gameplay. The model learned to predict future frame embeddings from current embeddings and actions — encoding billiards physics purely from pixels. **Probe results** (linear probe on encoder representations; 192-dim = pre-projector CLS token, 32-dim = post-projector): | Model | Rep dim | pos R² | vel R² | |---|---|---|---| | `lewm_small` | 32 | 0.983 | 0.296 | | `lewm_mamba` | 32 | 0.983 | 0.297 | | `lewm_framestacked` | 192 | 0.446 ⚠️ | 0.138 | | `lewm_framestacked` | 32 | 0.599 | 0.417 | | `lewm_auxloss_full` | 192 | **0.999** ✅ | **0.947** ✅ | | `lewm_auxloss_full` | 32 | 0.982 | 0.554 | > **Key finding:** Frame-stacking causes JEPA representational eviction — the ViT encoder stops encoding ball position (pos R²=0.446 at 192-dim) under optical-flow pressure. Adding a lightweight auxiliary state supervision head (λ=0.1) fully recovers position encoding (pos R²=0.999) and achieves the best prediction loss across all variants. ## Planning results | Approach | Same-episode | Novel cross-episode | |---|---|---| | Pure JEPA embedding CEM | ❌ FAIL | ❌ FAIL | | State-based hybrid C...

Note11. Julihuggingface.co/hellojais/lewm-billiards

Source context: 0 downloads · 0 likes · Repo hellojais/lewm-billiards

Note24. Maihuggingface.co/api/models/hellojais/lewm-billiards

Source context: 0 downloads · 0 likes · Repo hellojais/lewm-billiards