SFT datasets: how to design a supervised fine-tuning corpus

A pillar guide to supervised fine-tuning corpora: instruction→answer contracts, coverage, refusals, synthetic vs expert sources, pre-train gates, and eval leakage.