Lab

Start noisy, finish clean — batch size as a training dial

A large batch averages many gradient signals before each update. That is efficient late in training when you want smooth, stable convergence. But early in training, a large batch can be wasteful — the model changes so fast that samples from the same large batch are nearly correlated. Small batches inject noise that pushes the model to explore.

PR #163 schedules the batch size in three stages: small early for high-variance exploration, then growing twice as training progresses. The batch size triples from start to finish. This saves compute where noise is free and buys stability where precision counts.