pcbnew100_fastwam_scratch — FastWAM (full 30/30 MoT), from scratch, place_cube_new 100 traj

FastWAM (full non-decoupled 30/30 MoT) action-video model trained from scratch (resume: null = no LIBERO init) on the 100-trajectory "pick and place new" dataset (place_cube_new_lerobot_v21, 100 episodes; task "place the cube in the bowl").

• From scratch: resume: null.
• Model: full FastWAM (non-decoupled MoT), not the decoupled FlashWAM/FusedKV variant.
• Recipe: global batch 32, lr 1e-4, bf16, checkpoints every 5 epochs.
• Status: training completed — all 6 checkpoints present (steps 2445…14670 = epochs 5/10/15/20/25/30).

Checkpoints (checkpoints/weights/)

file	size
step_002445.pt	12.04 GB
step_004890.pt	12.04 GB
step_007335.pt	12.04 GB
step_009780.pt	12.04 GB
step_012225.pt	12.04 GB
step_014670.pt	12.04 GB

Weights-only checkpoints. Pair each with THIS run's dataset_stats.json for action/proprio (de)normalization.

Other files

• config.yaml — full training/model config.
• dataset_stats.json — per-run normalization stats.

Related

SleepMastger/pcbnew100_flashwam_scratch — same data, FlashWAM (FusedKV) variant, also from scratch.