pcbnew200_flashwam_scratch — FlashWAM (FusedKV / RoPE-fixed), from scratch, place_cube_new 200 traj

FlashWAM (M1_FusedKV_RopeFixed) action-video model trained from scratch (no LIBERO init) on the 200-trajectory "pick and place new" dataset.

• Model: FasterWAM decoupled, kv_source_mode: fused_kv, fixed_rope: true (video DiT Wan2.2-TI2V-5B backbone + 1-layer action DiT).
• Init: resume: null — trained from scratch (Wan2.2 base video expert, random action expert), NOT fine-tuned from a LIBERO checkpoint.
• Data: place_cube_new_lerobot_v21 — 200 episodes / 26,437 frames, 10 Hz, 2 cameras (base + wrist), 7-dim delta action, 8-dim proprio (real gripper widths). Task: "place the cube in the bowl" (pick-and-place-new).
• Training: 30 epochs / 24,810 steps, global batch 32, lr 1e-4 cosine, bf16. Completed cleanly (max_steps reached). Final loss=0.0501, loss_action=0.0062.

Checkpoints (checkpoints/weights/)

Weights-only checkpoints saved every 5 epochs:

file	step	epoch
step_004135.pt	4,135	5
step_008270.pt	8,270	10
step_012405.pt	12,405	15
step_016540.pt	16,540	20
step_020675.pt	20,675	25
step_024810.pt	24,810	30 (final)

Each file is ~10.13 GB (bf16 full model: fused video + action experts).

Other files

• config.yaml — full training/model config for this run.
• dataset_stats.json — per-run action/state normalization stats (min/max), needed at inference for de/normalization. Always pair a checkpoint with THIS run's stats.

Related

Companion run SleepMastger/random_new is the same FlashWAM architecture but fine-tuned from LIBERO on the 500-trajectory place_cube_new dataset.