AI 101

Fine-tuning

Also known as: SFT · RLHF

Training

Fine-tuning takes an already pre-trained model and gives it a second, lighter training round — this time on data chosen for behavior, not just language.

The usual recipe:

  1. Pre-trained model (knows language, not yet “how to be helpful”).
  2. Fine-tuning (a.k.a. SFT) — thousands of good example conversations: user asks, assistant answers well. The model learns the style of helpful.
  3. Preference training (often called RLHF) — the model is shown pairs of answers, humans mark which one is better, and it learns to prefer the better kind.

The result: a model that answers instead of just continuing the sentence.

Fine-tuning vs RAG — the classic confusion: fine-tuning changes the *brain* (slow, permanent, good for style and skills). RAG changes the *notes the model can consult* (instant, cheap, good for fresh or private facts). Beginners usually need RAG, not fine-tuning.