Group-LAM: completed Euclidean and RoTScale baselines
Private archive of the completed recon-frame-local-20260907-01 comparison. This repository contains the original, unmodified final checkpoints for both geometries and both training phases, together with their original runtime Hydra configurations and overrides.
These are prediction-only, frame-local baselines, not the newer group-loss experiment. Composition, inverse, and cross-step group objectives were disabled; the subsequently added inverse-state objective was not active.
Checkpoints
| Geometry | Phase | Training mode | Completed optimizer steps | File |
|---|---|---|---|---|
| Euclidean | 1 | Teacher-forced one-step | 15000 | RECON_Phase1_Euclidean_Prediction_final.pt |
| RoTScale | 1 | Teacher-forced one-step | 15000 | RECON_Phase1_RotScale_Prediction_final.pt |
| Euclidean | 2 | Autoregressive | 15000 | RECON_Phase2_Euclidean_Autoregressive_final.pt |
| RoTScale | 2 | Autoregressive | 15000 | RECON_Phase2_RotScale_Autoregressive_final.pt |
Every file was verified to have checkpoint_kind=resumable, steps=15000, next_step=15000, and model, optimizer, scheduler, checkpoint-contract, and provenance entries. P2 started from the matching geometry's completed P1 model with a fresh optimizer.
The four checkpoint files total 4,661,694,092 bytes. They retain the original PyTorch .pt format and are not standard Transformers from_pretrained models.
Configuration and integrity
Each {geometry}_phase{phase}_config.yaml and {geometry}_phase{phase}_overrides.yaml comes from the actual corresponding run's rank-0 Hydra output, not today's development configuration.
manifest.json records checkpoint sizes, SHA-256 hashes, Azure source paths/ETags, and the original training source digest. SHA256SUMS covers the checkpoint and runtime configuration files. Checkpoints were loaded for validation with torch.load(..., map_location="cpu", weights_only=True); optimizer and scheduler states were preserved.
Original source digest:
sha256:22ab7375e12a16e7114869f58e78bd42449d9276900ef5d027f26bd7a6d53d74
Download and use
Authenticate with an account authorized to access this private repository, then download a selected artifact, for example:
hf download tianqiuMS/Group-LAM-checkpoints RECON_Phase2_RotScale_Autoregressive_final.pt --local-dir ./checkpoints
Use the Group-LAM training/evaluation code and the archived matching configuration. Full resume must preserve the checkpoint's training contract; switching to newer group-loss defaults is not the same as resuming these prediction-only baselines. Machine-specific paths in the original configs need overrides on another host.
External RECON data, frozen DINOv3/RAE assets, and their source/dependencies are not included. This archive does not claim to be a self-contained training environment. Autoregressive actions in the original experiment were inferred from ground-truth frame pairs; these checkpoints are not command-conditioned controllers.
No datasets, access tokens, or unrelated experiment files are intentionally included. Keep this repository private unless an explicit separate decision authorizes broader distribution.