MolmoBot Img-DROID FrankaPickHard50 AuxFix LoRA 1k
This checkpoint is a 1,000-step MolmoBot finetune on
frankapickhard50_scripted_2demo_auxfix, starting from
allenai/MolmoBot-Img-DROID.
Important run constraints:
- Base checkpoint:
allenai/MolmoBot-Img-DROID
- Dataset:
ccwatson/frankapickhard50_scripted_2demo_auxfix
- Steps: 1,000
- Global batch size: 1
- Precision: AMP bf16
- Camera preset:
franka_wrist_only
- Action preset:
franka_joint
- Observation steps: 1
- Sequence length: 256
- LoRA: enabled on the final LLM block only
- LoRA rank: 4
- LoRA alpha: 8
- LoRA dropout: 0.05
- LoRA target preset:
attention
- Action expert: reset and trained with 4 layers
- FSDP: disabled for this single-GPU run
The artifact is a full unsharded MolmoBot checkpoint, not an adapter-only export.
It includes frozen base weights, trained LoRA tensors, and the trained reset
action expert.
W&B run:
https://wandb.ai/ccwatson/molmobot_lora_finetunes/runs/bsmggw34
