molmoact2_busybox_multitask
Full-component fine-tune of MolmoAct2 on SO-101 BusyBox multitask. Token embeddings stay frozen.
| Init | allenai/MolmoAct2 (not the SO100_101 checkpoint) |
| Dataset | armnet/busybox_multitask (132 episodes, 27 tasks) |
| Cameras | top, wrist, front |
| Training | 1× H200 NVL, batch 4, 20k steps, horizon 30, gradient checkpointing on |
| Result | loss 0.231 → 0.017. 5 h 21 m. Finished 2026-10-25 17:54 UTC. 5.046B learnable / 5.442B total |
| W&B | olalqzcn (offline run, partial sync) |
Checkpoints
checkpoint-20000/ is the final step.
checkpoint-5000/ · checkpoint-10000/ · checkpoint-15000/ · checkpoint-20000/
from huggingface_hub import snapshot_download
ckpt = snapshot_download("armnet/molmoact2_busybox_multitask", allow_patterns="checkpoint-20000/**")