SemWAM DINO S-VAE — LIBERO v2
Frozen semantic feature compressor for SemWAM. It compresses temporally pooled DINOv3 ViT-B/16 features from 768 to 48 channels while preserving the spatial and temporal grid. This is a standalone compressor, not a trained downstream WAM or robot policy.
Released October 25, 2026. The run completed normally at 150,000 optimizer steps; the selected best artifact records step 150,000.
Architecture and contract
Encoder and decoder: 3 transformer layers, 16 attention heads, FFN width 2048, dropout 0. Input dimension 768, latent dimension 48. Use deterministic posterior mean (encode_mean) for downstream features. Layout: libero_horizontal; input paradigm: t_concat.
The recommended downstream recipe is libero_svae/default, with parameter-free LayerNorm (eps=1e-6) on the compressed output and semantic velocity prediction. Output LayerNorm belongs to the downstream integration; it is not a replacement for the artifact's stored input normalization.
best.pt is the original SemWAM artifact, including model configuration, frozen weights, fitted input statistics, encoder/collection provenance and content identity. It must be loaded using SemWAM's artifact loader, not treated as a plain state dictionary. Use the matching DINO encoder, camera layout and sampling geometry; preserve all identity checks. This compressor is specific to the LIBERO release and split used in pretraining.
Compression evaluation
Held-out reconstruction audit: 6,846 validation windows, bf16 inference in standardized DINO feature space.
| Metric | S-VAE | PCA-48 |
|---|---|---|
| Standardized MSE | 0.042676 | 0.328235 |
| Explained variance | 95.74% | 67.21% |
| Mean token cosine | 0.978378 | 0.818406 |
No dead latent channels were found. Full reconstruction and latent health, generalization, similarity and precision results are included in the JSON reports. PCA is fitted only on sampled training features.
Wrist-camera and first-frame reconstruction errors are higher than main-camera and pooled-frame errors.
Compression quality and loader checks support trying this artifact downstream. They do not establish downstream convergence, action quality, closed-loop success or superiority to raw DINO features. The published artifact is not a downstream policy checkpoint.
Use with SemWAM
Source repository: https://github.com/RainyRobo/SemWAM
from huggingface_hub import hf_hub_download
from wam.models.semantic.svae.artifacts import load_artifact
path = hf_hub_download("RainyBot/SemWAM-SVAE-LIBERO-v2", "best.pt")
compressor, metadata = load_artifact(path)
compressor.eval().requires_grad_(False)
# z = compressor.encode_mean(dino_features)
For downstream configuration, set DINO_SVAE_WEIGHTS to the downloaded artifact path and select libero_svae/default. Model, data and source contracts are validated by SemWAM.
Provenance and files
best.pt: original complete compressor artifact.model_config.json: compressor architecture.pretraining_config.yaml: standalone pretraining recipe.artifact_identity.json: encoder, split, collection and normalization provenance.verification.json: SHA256 hashes and downstream contract.reconstruction_quality.json,latent_usability.json: held-out compression audits.
State SHA256: c6c26c8b5883a6360d0849bf864c9ba2fe0056f5b9c921860ecc29bfe6e9e05b.
File SHA256: 70bed2697f9e5ea878fbaa5ba050a73d9f26220fb52da623d6c86debb9e3b1b4.
DINOv3 weights, datasets, feature caches, optimizer states, credentials and downstream WAM weights are not bundled. Consult the original DINOv3 and dataset terms for those dependencies; no new license grant is made by this model card.