| --- |
| library_name: lerobot |
| tags: |
| - molmoact2 |
| - robotics |
| - lerobot |
| - vla |
| - quantile-normalization |
| base_model: allenai/MolmoAct2-SO100_101 |
| --- |
| |
| # molmoact2_block_stack_so101_quantile_12k |
| |
| Fine-tuned [MolmoAct2](https://huggingface.co/allenai/MolmoAct2) (action-expert-only) on block_stack with **QUANTILES** normalization (q01/q99). Intermediate checkpoint at step 12000 while 30k training continues on Isambard. |
|
|
| | | | |
| |---|---| |
| | **Policy** | MolmoAct2 (`policy.type=molmoact2`) | |
| | **Init checkpoint** | [allenai/MolmoAct2-SO100_101](https://huggingface.co/allenai/MolmoAct2-SO100_101) | |
| | **Dataset** | [villekuosmanen/armnetbench_block_stack](https://huggingface.co/datasets/villekuosmanen/armnetbench_block_stack) | |
| | **Task** | `block_stack` | |
| | **Local run** | `molmoact2_block_stack_so101_quantile` | |
| | **Checkpoint step** | `012000` (12k / 30k target) | |
| | **Normalization** | QUANTILES (action + state + gripper), IDENTITY (visual) | |
| | **Training** | Isambard GH200, batch 64 (16/GPU x 4 DDP), bf16, no gradient checkpointing | |
|
|
| ## Checkpoints |
|
|
| Step `012000` lives at the **repository root** for direct loading. |
|
|
| ## Usage |
|
|
| ```python |
| from lerobot.policies.molmoact2.modeling_molmoact2 import MolmoAct2Policy |
| policy = MolmoAct2Policy.from_pretrained("pravsels/molmoact2_block_stack_so101_quantile_12k") |
| ``` |
|
|
