smolvla_makermods_arrange_chocolate_finetune_2026-08-28_19-27-10
25,000-step SmolVLA trained from lerobot/smolvla_base, on makermods/arrange_chocolate_finetune.
Training
| Policy | SmolVLA (smolvla) |
| Base | lerobot/smolvla_base |
| Dataset | makermods/arrange_chocolate_finetune |
| Steps | 25000 |
| Batch size | 64 |
| LR | 0.0001 -> 2.5e-06 (cosine_decay_with_warmup) |
| Warmup steps | 1000 |
| Cameras | front, top, wrist (3x 480x640) |
| Checkpoints | 25 (every 1000 steps) |
Layout
The final checkpoint is at the repo root, so this repo loads directly as a policy.
Every intermediate checkpoint is under checkpoints/<step>/, each with
pretrained_model/ and training_state/ (resumable).
rename_map is empty and the camera keys are the real dataset keys
(observation.images.front / .top / .wrist), so lerobot-eval,
lerobot-rollout and policy_server.py load this checkpoint without a rewrite.