Thank you for your great work! I want to know how to save model and evaluate on downstream LLM tasks reported on Megalodon paper: 