Training on large quantities of LLM-generated synthetic data is an important part of training LLMs.