molmoact2_busybox_multitask_so101

Full-component fine-tune of MolmoAct2-SO100_101 on SO-101 BusyBox multitask. Token embeddings stay frozen. This card is the 4× H200 run at effective batch 32, with image augmentation off.

Init	allenai/MolmoAct2-SO100_101 @ 152569fe
Dataset	armnet/busybox_multitask (132 episodes / 25,734 frames, 27 tasks)
Cameras	top, wrist, front
Training	4× H200, effective batch 32 (8/GPU), 20k steps, horizon 30, gradient checkpointing off, bf16
Result	loss 0.334 → 0.002. 7 h 21 m. Finished 2026-10-28 04:44 UTC. 5.046B learnable / 5.442B total
W&B	03krz2v5

Checkpoints

checkpoint-20000/ is the final step. Tags step-5000, step-10000, step-15000, and step-20000 point at this run.

from huggingface_hub import snapshot_download

ckpt = snapshot_download(
    "armnet/molmoact2_busybox_multitask_so101",
    revision="step-20000",
    allow_patterns="checkpoint-20000/**",
)