nth-lr-rel10-ie-mfe-best-20ksteps

MolmoBot relabeled directional next-to left/right — step 20,000

This repository contains the unsharded MolmoBot checkpoint from training step 20,000 of nth-lr-rel10-ie-mfe-best-30k-b16-32.

It was uploaded separately while training was still running because it had the best intermediate closed-loop success rate observed through step 24,000. This repository is a fixed snapshot of step 20,000; it will not be replaced by later checkpoints from the run.

Files

• model.pt: unsharded MolmoBot model weights
• config.yaml: effective MolmoBot training/model configuration
• intermediate_eval_step20000.json: compact closed-loop evaluation metadata

The checkpoint directory is directly usable by MolmoBot inference tooling and the vla-evaluation-harness MolmoBot model server.

Training

• Base model: allenai/MolmoBot-DROID
• Dataset: ccwatson/RelabeledDirectionalPnPNextTo-10shard-left-right-only
• Checkpoint step: 20,000
• Train/validation trajectories: 2,903 / 37
• Directions: left and right
• Action preset: franka_joint
• Camera preset: franka_droid
• Device/global batch size: 16 / 32
• Sequence length: 928
• Observation frames: 2, separated by 8 dataset steps
• Image augmentation: enabled
• Training task-description prompt randomization: disabled

Intermediate closed-loop evaluation

The checkpoint was evaluated on the 20-episode LR_10each unicorn benchmark: ten left episodes followed by ten right episodes.

• Success rate: 25% (5/20)
• Lax success rate: 25% (5/20)
• Coverage: 20/20 episodes
• Missing or duplicate episodes: 0
• Evaluation errors: 0
• Failed shards: 0
• Renderer: classic
• Parallel shards: 10
• Task horizon: 500 simulator steps

This is a small checkpoint-selection benchmark, not a final estimate over the full 81-left and 81-right benchmark suites.

Use

Download the repository to a local checkpoint directory:

hf download ccwatson/nth-lr-rel10-ie-mfe-best-20ksteps \
  --local-dir /path/to/nth-lr-rel10-ie-mfe-best-20ksteps

Configure the MolmoBot model server with that directory as its checkpoint path. A valid local directory contains both config.yaml and model.pt.

Run provenance

• Training W&B run: https://wandb.ai/ccwatson/molmobot_finetune/runs/nhmj55qz
• Intermediate-eval W&B run: https://wandb.ai/ccwatson/molmobot_finetune/runs/external-eval-nth-lr-rel10-ie-mfe-best-30k-b16-32-step20000
• Dataset revision checked before training: 92a2fd1aed2cb6ae5e7af3963c0f2b1f19110738
• Upload date: July 24, 2026