PixelWAM checkpoints

This repository contains final raw checkpoints, their matching DeepSpeed resume states, and the corresponding completed evaluation results for FastWAM / PixelWAM experiments.

Only the final checkpoint of each completed run is published. Intermediate checkpoints, fastwam_release checkpoints, and EMA checkpoints are excluded.

Layout

libero/
  pixel-asymflow/<training-recipe>/
  latent-fastwam/<training-recipe>/
robotwin/
  latent-fastwam/<training-recipe>/

Each run directory can contain:

• checkpoint/: final raw inference checkpoint.
• deepspeed_state/: matching optimizer, scheduler, RNG, and model resume state.
• evaluation/: completed benchmark output, including per-task JSON results, resolved configs, and summary.json when available.
• metadata/: historical resolved training config and dataset statistics.

See CHECKPOINT_INDEX.md for the complete mapping and evaluation status.