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

Downloads last month: 5

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support