nth-lr-vpconcat-ie-mfe-best-26k
Best checkpoint from the Betty MolmoBot run
nth-lr-vpconcat-ie-mfe-best-30k-b16-32, finetuned from
allenai/MolmoBot-DROID on the vanilla-procedural concat next-to dataset
(generic plus direction-conditioned language).
Selection and final evaluation:
- intermediate selection benchmark: 55% success (11/20), the highest result across all evaluated training checkpoints
- held-out left benchmark: 49.4% success (40/81)
- held-out right benchmark: 54.3% success (44/81)
- final step-30k checkpoint scored 46.9% left and 48.1% right; this step-26k checkpoint is superior on both
Training settings:
- steps at export: 26,000
- 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
- checkpoint format: unsharded MolmoBot (
config.yaml, model.pt)
external_eval_request.json records the intermediate evaluation request.
final_best_checkpoint_external_eval_request_*.json records the two held-out final evaluation requests.
