dddddsds31341/Long-WAM-Four-Tasks-Step-40000

Long-WAM Four Tasks, step 40000

Checkpoint from piper-four-tasks-50k-20261109, continued from step 25000 toward 80000. The saved optimizer step is 40000.

• model.pt: original inference weights, including the joint model and proprioception encoder.
• training_state/step_040000/: original four-rank DeepSpeed/Accelerate model, optimizer, scheduler, RNG and trainer/sampler state, including the completion marker.
• training_config.yaml and dataset_stats.json: resolved training configuration and the original state/action normalization statistics.
• code.zip: the existing long-wam-sip-four-tasks-80k-resume25k-save10k-20261111 code release.

Inference must use the supplied code, task configuration and dataset_stats.json. Required pretrained components and datasets are referenced by training_config.yaml and are not duplicated here.

For continuation, restore the full training_state/step_040000 directory, including hidden files, and override resume with its restored absolute path when launching the recorded four-rank training configuration. Adjust machine-specific paths. The archived config's original resume path names 25k; use the restored 40k path instead.

Training completion does not establish robot success rates.