After this lesson, you will be able to: Map the pretraining to supervised-fine-tuning to alignment lifecycle, and decide between prompting, retrieval, and fine-tuning for a given problem.
Fine-tuning is the research skill that turns a general base model into something specific. But it is often the wrong tool. This intro frames the adaptation lifecycle and the honest decision of when to fine-tune at all.
This is a free introductory lesson. No purchase required.
A usable assistant is built in stages. Pretraining produces a base model that completes text. Supervised fine-tuning (SFT) on instruction/response pairs teaches it to follow instructions. Preference optimization (RLHF or DPO) aligns it with human preferences for helpfulness and harmlessness. Each stage uses less data than the last and changes the model's behavior more than its raw knowledge. As a researcher you will almost always start from a pretrained (and often instruction-tuned) open model and run one of the later stages, not pretrain from scratch.
Good at: format and style (always output JSON, always answer in this tone), narrow task performance (classification, extraction, a specific generation task), and distilling a large model's behavior into a small one. Bad at: injecting large amounts of new factual knowledge reliably (it tends to learn the form of answers, not the facts, and can increase hallucination if you fine-tune on facts it half-learns). Knowing this boundary is exactly what separates a sound experimental design from a doomed one.