Ollama 3.1 405B model already uses synthetic data for SFT. They basically trained a bunch of models to generate, correct, and reject synthetic data for various tasks. They cover it in section 4.3 in the "The Llama 3 Herd of Models" paper.