Where A,B,C are empirical constants,

Loss = A/parameters + B/dataset_size + C

Just read the chinchilla paper.