Tag
Proposes a three-term scaling law that decouples model size, training steps, and batch size, enabling robust fitting with fewer runs and deriving scaling laws for suboptimal batch sizes.