SmolVLA · blue cube → orange box (SO-101)

Training

• 20,000 steps, batch size 64 (~23 epochs), bf16 AMP, RTX 4090
• lr 1e-4, cosine decay to 2.5e-6 over 20k steps, 1k warmup
• Frozen vision encoder, action expert only (train_expert_only=true)
• Final smoothed loss ≈ 0.045
• Full config in train_config.json

Inference note