Fine-Tuning: Adapting Giants

A modern LLM ships as a frozen, general-purpose giant. Making it yours is a craft with its own math and its own traps: adapting billions of parameters by training a few million (LoRA), aligning behaviour instead of teaching facts (SFT, RLHF, DPO), and the production realities that sink naive fine-tunes — catastrophic forgetting, eval gates, and the constant question of whether you should be fine-tuning at all. Studied the same way as everything else: from the problem up, mocked at interview intensity, then built in public.