SemWAM RoboTwin Clean2Random v1
SemWAM checkpoint for RoboTwin. This repository contains the final model weights and the configuration and normalization statistics needed for inference.
Checkpoint
- Version: v1
- Optimizer step: 32,220
- Checkpoint format: v2
- Weight file:
checkpoints/weights/step_032220.pt - All original training launchers completed successfully.
The selected checkpoint, saved model contract, normalization statistics and tensor finite values were verified before publication. Publication does not establish closed-loop benchmark performance.
Files
| File | Purpose |
|---|---|
checkpoints/weights/step_032220.pt |
Model weights, including the frozen semantic compressor |
inference_config.json |
Saved inference contract |
config.yaml |
Inference configuration reconstructed from the saved contract |
dataset_stats.json |
Matching action and state normalization statistics |
checkpoint_manifest.json |
File identities and checksums |
model_card_summary.json |
Machine-readable model summary |
trainer_state.reference.json |
Step and inference reference; not a resumable training state |
Restoration
Use the SemWAM runtime to restore the saved v2 contract before constructing the model and loading its weights. Use the supplied normalization statistics and external frozen assets matching the saved identities. Required external asset hashes are listed in checkpoint_manifest.json.
Training data, private run configuration, initialization provenance, augmentation recipe, logs, optimizer, scheduler and per-rank RNG states are not included.
Integrity
- Weight SHA256:
a56f3e98e5b21ca44106c6382f1b589641f03f8badd572a804ba38e5e72e3708 - Statistics SHA256:
822d7bb9011031653fe8c92c2b716a4f808e93ba16da2e23a48b9eca2eca7a96 - Package verification time (UTC):
2026-11-08T11:37:47.628712+00:00
- Downloads last month
- 21