Fine-tuning is taking an already-trained model and further training it on your examples to tune the behavior for your task. The model already "knows the language" from base training; you tune it on your data so it answers in the needed style, format or for a specific task. This is one customization technique but, as we will see, not always the best.
The idea of fine-tuning
Take a ready model: it is already trained on huge data, knows language and reasoning.
Tune it on your examples: a set of "input to desired output" from your task.
The model tunes: it starts behaving closer to your examples.
What fine-tuning does well
What does fine-tuning really give, and what not?
It is important to understand what fine-tuning is strong for and what not — this saves you from a common mistake. Fine-tuning is GOOD for tuning BEHAVIOR: 1) Style and tone: so the model answers in your brand style, a certain voice, form. 2) Format: if you need a stably specific output format hard to describe with a prompt. 3) A specific task/pattern: teach the model a concrete type of task on many examples (e.g. a special kind of classification, transformation). 4) Prompt brevity: what you would have to explain each time with a long instruction can be "baked" into the model by tuning. Fine-tuning is BAD for adding KNOWLEDGE/FACTS: 1) "Remember our documents" is a common misconception. Fine-tuning remembers facts poorly, the model can still invent, and updating knowledge is costly (retraining on every change). For facts over your data you need RAG, not fine-tuning (we covered this in the RAG course). 2) Fresh/changing data — fine-tuning is static, retraining for changes is costly. The key distinction: fine-tuning changes HOW the model answers (behavior, style, format), not WHAT it knows (facts). For knowledge — RAG or context; for behavior — fine-tuning. A common costly beginner mistake: trying to fine-tune a model so it "knows our data". Usually the wrong tool — costly, complex, and it still invents. Correct: RAG for knowledge, fine-tuning for style/behavior/format, and even then — when prompting cannot cope. The whole next module is about this.
Fine-tuning changes "how", not "what"
A key understanding of the whole course: fine-tuning tunes how the model answers (style, tone, format, task type), not what it knows (facts). To add knowledge over your data you need RAG or context, not fine-tuning. Confusion here is a source of costly mistakes: people try to "fine-tune so it knows our documents", and it does not work. Keep this distinction in mind the whole course.
Course rule: fine-tuning is training a ready model on your examples to tune behavior (style, format, task). It changes "how" the model answers, not "what" it knows — for knowledge you need RAG.
Fine-tuning is training a ready model on your examples to tune its behavior.