pnp-lr-proc-ie-mfe-4kstep

MolmoBot finetune from allenai/MolmoBot-DROID on ccwatson/PnPNext-To-lr-pairs-procedural.

This is the current best checkpoint from the Betty training run pnp-lr-proc-ie-mfe-best2-30k-b16-32, selected at step 4,000 by its intermediate closed-loop evaluation.

Intermediate evaluation:

• benchmark: FrankaPickAndPlaceRelativeNextToProceduralPairsValSubsample10Pairs3HousesBench
• episodes: 20
• success rate: 45% (9/20)
• evaluation completed: 2026-08-07

Training settings:

• steps at export: 4,000
• device batch size: 16
• global batch size: 32
• action preset: franka_joint
• camera preset: franka_droid
• sequence length: 928
• observation steps: 2
• observation step delta: 8
• max images: 4
• trainable LLM: true
• checkpoint format: unsharded MolmoBot (config.yaml, model.pt)

external_eval_request.json preserves the evaluation request/provenance for this checkpoint.