You might want to give a read to "Scaling Data-Constrained Language Models" [1]. They basically generalized the Chinchilla scaling law by investigating behavior on multi-epoch runs.

[1] https://arxiv.org/abs/2305.16264