JHU

A large-scale empirical invesigation of scaling laws shows that performance has a power-law relationship to model size, dataset size, and training compute, while architectural details have minimal effects.