S016 Blocks Ranking Sequential DP — step 141394 training state
This is an intermediate recovery checkpoint for CTR Experiment E771, not a completed training or evaluated model. It is published to resume training after a GPU change.
- Source: E771-R001, saved at optimizer update 141394.
- Target budget: 424177 updates (600 eligible-window epochs), effective batch 128, seed 87431.
- Dataset: Shiki42/ctr-sortblocks-100ep-sequential, revision 3d8db515f856c4e6c2612569b43fb15eeaef3064.
- Official LeRobot 0.4.4 source: 8fff0fde7c79f23a93d845d1a50e985de01f8b8a.
- Policy: Diffusion Policy; observation horizon 2, prediction horizon 16, action queue 8; IdleMask disabled.
- Numerical normalization: source checkpoint MIN_MAX state/action and ImageNet image statistics, unchanged.
Contents and resume
The native pretrained_model/ directory includes model weights, resolved config and both processor states. training_state/ includes optimizer state and parameter groups, scheduler state, RNG state and the step counter. Verify every file using SHA256SUMS before use. Resume using the original LeRobot recipe and update machine-specific paths without changing the training budget or normalization. Native LeRobot resume restores model/optimizer/scheduler/RNG, but does not restore the exact data-iterator position.
The source Run failed while saving step 212091 because the disk filled; that incomplete checkpoint is not included. E771-R002 later resumed from this checkpoint, then was stopped by the user for GPU migration; its 963 unsaved updates were explicitly discarded. The published checkpoint remains step 141394.
CPU strict model and full optimizer/scheduler reload, finite-state checks, step validation and file hashes passed. No success-rate or completed-training claim is made; audit pending.