nt-postcond-mix30-30-40-100k-b16-32-6k

MolmoBot relative placement and orientation — 6k checkpoint

This is the 6,000-step unsharded checkpoint from nt-postcond-mix30-30-40-100k-b16-32, a MolmoBot-DROID fine-tuning experiment targeting 100,000 training steps. The linked W&B run contains the training history and logged evaluations.

Provenance

• Base model: allenai/MolmoBot-DROID.
• Source training run: nt-postcond-mix30-30-40-100k-b16-32.
• W&B run ID: uuaqppy6.
• Source Slurm job: 8231747.
• Checkpoint: step6000-unsharded, exported September 9, 2026.

Training setup

Training uses a 30/30/40 sampling mixture, independent of dataset sizes:

Prompts use sampled object names and natural-language postconditions for relative-placement tasks (next_to_with_natlang_postconditions). Point prompting is disabled. Device/global batch sizes are 16/32. Action and camera presets are franka_joint and franka_droid, with two observation frames separated by eight steps and an action horizon of 16.

Files and use

• model.pt: unsharded model weights.
• config.yaml: model/training configuration, including action and state normalization.

Load these files with a MolmoBot-compatible inference implementation, such as the MolmoBot server in the VLA evaluation harness. This repository does not contain optimizer or trainer state for a full-state training resume.

This is an intermediate checkpoint, not the completed 100k-step model. The experiment's 23-episode iTHOR benchmark is a smoke evaluation overlapping source demonstrations; its results should not be interpreted as held-out generalization.