Item Assembly Alignment Specialist B
Frozen specialist checkpoint selected for the item_assembly alignment benchmark.
Benchmark evidence
- Checkpoint-selection benchmark, full task, official
randomreset, 100 episodes. - B: 63/100; Original SmolVLA A: 53/100; B−A: +10 percentage points.
- This is a checkpoint-selection benchmark, not final competition validation.
- A and B used the same numeric seed population as a population-level control. The runs are not claimed to reproduce identical physical simulator states.
Checkpoint
- Source:
experiments/item_assembly_alignment_specialist/checkpoint_sprint_v1/training/selected_checkpoint_B.pt - SHA256:
9b5ba47b8c94ba5a03f6721ada2ed335a41eedaec0f658482e71670938a9730a - File:
selected_checkpoint_B.pt