What was the most frustrating part of your last local fine-tune?

I’m working on a local fine-tuning tool, and I’m curious where people actually lose the most time. Was it getting the environment working, preparing the dataset, fitting everything into VRAM, or getting the exported...