---
tags:
- molmoact2
- robotics
- lerobot
- vla
datasets:
- armnet/busybox_multitask
base_model: allenai/MolmoAct2
---

# molmoact2_busybox_multitask

Full-component fine-tune of [MolmoAct2](https://huggingface.co/allenai/MolmoAct2) on SO-101 BusyBox multitask. Token embeddings stay frozen. This card is the 4× H200 retrain at effective batch 32.

| | |
|---|---|
| **Init** | `allenai/MolmoAct2` (not the SO100_101 checkpoint) |
| **Dataset** | [armnet/busybox_multitask](https://huggingface.co/datasets/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 0.220 → 0.005. 14 h 23 m. Finished 2026-10-27 12:00 UTC. 5.046B learnable / 5.442B total |
| **W&B** | [dhseedj1](https://wandb.ai/pravsels/molmoact2_busybox_multitask/runs/dhseedj1) |

## Checkpoints

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

```python
from huggingface_hub import snapshot_download

ckpt = snapshot_download(
    "armnet/molmoact2_busybox_multitask",
    revision="step-20000",
    allow_patterns="checkpoint-20000/**",
)
```

## Previous run

The 2026-10-25 1× H200 batch-4 fine-tune ([olalqzcn](https://wandb.ai/pravsels/molmoact2_busybox_multitask/runs/olalqzcn)) is still on this repo at `olalqzcn-step-5000`, `olalqzcn-step-10000`, `olalqzcn-step-15000`, and `olalqzcn-step-20000`. Load that checkpoint with `revision="olalqzcn-step-20000"`.