pcbnew_flashwam_ft — FusedKV / RoPE-fixed FlashWAM, finetuned on place_cube_new (500 traj)
FlashWAM (M1_FusedKV_RopeFixed) action-video model, finetuned on the new
500-episode place-cube-in-bowl dataset ("pick and place new").
- Model: FasterWAM decoupled,
kv_source_mode: fused_kv,fixed_rope: true(video DiT Wan2.2-TI2V-5B backbone + 1-layer action DiT). - Data:
place_cube_new_lerobot_v21— 500 episodes / 51,579 frames, 10 Hz, 2 cameras (base + wrist), 7-dim delta action, 8-dim proprio (real gripper widths). Task string: "place the cube in the bowl". - Init: finetuned from LIBERO checkpoint
M1_FusedKV_RopeFixed_step021700(new fused_kv format). - Training: 30 epochs / 48,360 steps, 4×H200, global batch 32, lr 1e-4 cosine,
bf16. Final
loss=0.0802,loss_action=0.0045.
Checkpoints (checkpoints/weights/)
Weights-only checkpoints saved every 5 epochs:
| file | step | epoch |
|---|---|---|
step_008060.pt |
8,060 | 5 |
step_016120.pt |
16,120 | 10 |
step_024180.pt |
24,180 | 15 |
step_032240.pt |
32,240 | 20 |
step_040300.pt |
40,300 | 25 |
step_048360.pt |
48,360 | 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.
- Downloads last month
- 18