A multi-scale, multi-fidelity Bayesian Optimization (BO) approach where {data mixtures, model scale, training steps} are adaptively selected, achieving >2.6x speedups compared to multi-fidelity BO and random search baselines.