What Does Fine-Tuning Actually Learn? Testing QLoRA with Deliberately Controlled Data

A controlled QLoRA experiment: format is learned first, facts come later, and brand drift turns out to just be about epoch count.