Downloads · 30 days
0
ggkiller-air/univlat-desk-sweep-training-artifacts
univlat-desk-sweep-training-artifacts is a robotics model from ggkiller-air. Use it for the robotics task on the model card, and read the license before you ship it in a product.
This repository archives the complete retained training outputs for the SONIC desksweep / sonichtd comparison run. It contains resumable checkpoints, optimizer and scheduler state, RNG state, best/final model exports,…
Downloads · 30 days
0
Access
Public
Updated Aug 16, 2026
Repo size
—
Likes
0
Public
Click a slice to open those files.
.json33.8 MB · 71%
From the Hugging Face model README
This repository archives the complete retained training outputs for the SONIC
desk_sweep / sonic_htd comparison run. It contains resumable checkpoints,
optimizer and scheduler state, RNG state, best/final model exports, resolved
configuration, and retained training logs.
| Backbone | Training result | Selected checkpoint |
|---|---|---|
| Isaac-GR00T | 50,000 steps | final 50k; validation was disabled, so no best checkpoint exists |
| Diffusion Policy | 25,000 steps | best 25k, validation action MSE 14.117382 |
| openpi pi0.5 | 10,000 steps | best 10k, validation action MSE 0.072074 |
| starVLA | 45,000 steps | best 35k, validation action MSE 0.004136 |
DiT4DiT is intentionally absent. Its ZeRO-3 run failed during its first validation and did not produce a usable resumable checkpoint.
Isaac-GR00T/
diffusion_policy/
openpi/
starVLA/
training_logs/
docs/
The logical archive size is approximately 498 GB. Intermediate checkpoints are kept deliberately so training can be resumed and checkpoint-selection claims can be audited. Raw training data is not included.
best means the lowest held-out action MSE among validation points for that
trainer only. Action normalization and validation implementations differ, so
the MSE values must not be compared numerically across backbones. Final policy
ranking requires common closed-loop evaluation.