szcharlesji/yor-patch-policy-checkpoints

YOR Patch Policy checkpoints

Private archive of the TRAINED YOR checkpoints and Thor deployment bundles held locally at backup time. Includes CLS and spatial-patch variants, camera/norm/chunk ablations, cup picking and door opening. Snapshots from active runs are pinned to their saved inode; they do not represent unsaved in-memory training progress.

This archive is deliberately incomplete. To fit a constrained uplink, untrained initializations (model_init.pt) and mid-training exports (model_49.pt, model_epoch199.pt, model_stopped.pt, snapshot_before_wandb_restore.pt) were NOT uploaded and exist only on the original machine. Every uploaded run keeps its snapshot.pt and/or model_final.pt. Runs whose only file was an untrained init are listed under skipped_runs in RUN_INDEX.json and are absent here.

Download

from huggingface_hub import snapshot_download
snapshot_download("szcharlesji/yor-patch-policy-checkpoints", local_dir="yor-checkpoints")

Use allow_patterns to download only the desired training_runs/<run>/ or thor_bundles/<bundle>/ files. Authenticate first. Verify downloads with SHA256SUMS (or the selected entries in upload_manifest.json).

Contents and use

• training_runs/: the trained .pt files per run and its exact Hydra configuration. snapshot.pt contains model, optimizer and resume metadata. model_final.pt is the final saved model. Per-run held_back in RUN_INDEX.json names the files left on the original machine.
• thor_bundles/: named inference bundles with policy.pt, encoder.pt, card.yaml and package-local runtime definitions/instructions.
• legacy_exports/: older run-local checkpoint export directories.
• code/: reconstruction code and recipes. Install dependencies from conda_env.yml; make code importable when loading whole-module PyTorch checkpoints.
• RUN_INDEX.json: uploaded_runs (architecture, dataset, saved epochs, checkpoint paths, held_back), skipped_runs, and excluded_weight_filenames. Filenames identify the original runs rather than inventing new training provenance.

Whole-module .pt files require their defining Python modules and torch.load(..., weights_only=False); load only this trusted private archive. VQ-BeT normalized models output raw actions. Filtered/jittered policies predict variable-time waypoints; inspect the card and README before choosing playback timing. Embedding caches and W&B credentials/logs are not included.

Saved runs

Run	Encoder	Views	Norm	Chunk	Batch	Saved epochs / planned
2026-09-04/00-53-42	x_norm_patchtokens	None	False	5	4	400 / 400
2026-09-08/16-24-49	x_norm_patchtokens	None	False	5	16	112 / 200
C-V1-M/2026-09-10/00-08-46	x_norm_clstoken	1	True	5	512	1000 / 1000
C-V1-R/2026-09-10/00-13-21	x_norm_clstoken	1	False	5	512	1000 / 1000
C-V3-M/2026-09-10/00-17-54	x_norm_clstoken	3	True	5	512	1000 / 1000
C-V3-M-C16-temporal-s3-filterFalse/2026-09-12/13-15-42	x_norm_clstoken	3	True	16	512	1249 / 2000
C-V3-M-C16-temporal-s3-filterTrue/2026-09-12/13-15-13	x_norm_clstoken	3	True	16	512	1413 / 2000
C-V3-M-C20/2026-09-11/15-34-47	x_norm_clstoken	3	True	20	512	1000 / 1000
C-V3-M-C20-temporal-s1-filterTrue/2026-09-12/11-21-21	x_norm_clstoken	3	True	20	512	2000 / 2000
C-V3-M-jittered-C8-next1/2026-09-12/20-01-47-seattle	x_norm_clstoken	3	True	8	512	2000 / 2000
C-V3-M-jittered-C8-next1/2026-09-12/22-44-00-leo-b64	x_norm_clstoken	3	True	8	64	2000 / 2000
C-V3-R/2026-09-10/00-08-46	x_norm_clstoken	3	False	5	512	1000 / 1000
P-V1-M/2026-09-10/00-32-12	x_norm_patchtokens	1	True	5	16	200 / 200
P-V3-M/2026-09-10/00-44-03	x_norm_patchtokens	3	True	5	4	51 / 200
P-V3-R/2026-09-10/00-45-01	x_norm_patchtokens	3	False	5	4	7 / 200
cup_patch_filtered_jittered_20260913_144151	x_norm_patchtokens	3	True	8	4	100 / 2000
door_cls_filtered_jittered_20260913_032844	x_norm_clstoken	3	True	8	64	2000 / 2000
door_cls_unfiltered_20260913_032844	x_norm_clstoken	3	True	8	64	2000 / 2000