Fine-tuning, SFT and where instruction-following comes from

In 2022, an ICT teacher of fifteen years tried a language model for the first time. He typed "Please explain photosynthesis" and expected an explanation. Instead the model continued: "Please explain respiration. Please explain transpiration…" He said in the staff room: "It thinks I'm setting a test." That reaction marks where this lesson sits in module 2: pre-training teaches what language looks like; fine-tuning teaches what to do when someone asks.

As lesson 6 explained, the only task in pre-training is guessing the next word. In text scraped from the web, the thing most likely to follow "Please explain photosynthesis" really is another question, not an answer. The model was not incapable of explaining. It did not know you wanted an explanation. SFT (supervised fine-tuning) closes that gap.

---

Read every episode, never hosted one

After pre-training, a model is like someone who has read every transcript of a phone-in radio show: fluent, well-informed, familiar with the catchphrases, but never held the microphone. Tell him to "take this listener call" and he may recite an opening monologue from a random episode instead of responding to the caller.

How does a station turn that person into a host? Not with ten more books, but by playing hundreds of demonstration call-ins: what the listener asked, how the host followed up, how the point was made in thirty seconds, how a complaint was acknowledged before replying. Enough demonstrations, and the habit forms — this kind of input gets this kind of output.

That is supervised fine-tuning, universally abbreviated to SFT.

What demonstration exchanges look like

The team assembles a large set of pairs: an instruction plus an ideal human-written response.

• Instruction: "Explain photosynthesis to a primary pupil in three sentences." Response: an explanation that really is three sentences and really is suitable for a primary pupil.
• Instruction: "Make this letter more formal." Response: an actually rewritten letter.
• Instruction: "What is your name?" Response: an introduction consistent with the product's positioning.

Training reuses the mechanism from lesson 6, except what gets covered up is the ideal response. Every wrong guess nudges the parameters; after hundreds of thousands of rounds the model has the habit of carrying out instructions rather than extending them — which is why, after 2023, it finally felt as though the model "answered questions".

Small volume, brutal quality bar

Pre-training runs on trillions of tokens. SFT typically uses tens to hundreds of thousands of demonstrations — several orders of magnitude smaller. Every item is written or reviewed by a person, which makes it expensive.

The reason is direct: at this stage the model is learning habits, and habits are easily corrupted by a small number of poor demonstrations. If one demonstration in ten is perfunctory or inconsistently formatted, the model learns that being perfunctory is sometimes acceptable — exactly like training a new host; cut a few corners in the demos and the habit copies.

Behaviour, not knowledge

Many panel heads' first instinct: fine-tune on all the school's curriculum documents so the AI "knows our school". It usually does not work. The model sees those documents during fine-tuning but learns what documents of that kind look like, not what page three says.

School facts still belong in a school knowledge base — citable, updatable, effective the moment a document is revised. Changing facts in a fine-tuned model means training again.

Fine-tuning teaches it how to say things; a knowledge base gives it what to say — worth pinning beside the panel meeting whiteboard.

Eighty per cent prompts, twenty per cent fine-tune

A secondary school wanted AI-drafted parent circulars to match the house format: fixed heading, paragraph order, customary courtesies, reply slip at the end. After two months of prompt templates the results were usable but inconsistent — every colleague wrote a slightly different prompt, and the template had grown past two hundred characters, which new staff simply would not use.

Step one, exhaust prompts: template into a shared folder, one colleague maintains it, must-include items become a checklist. About eighty per cent closed, zero cost.

Step two, fine-tune the remainder: a light fine-tune on the on-premises model using two hundred real circulars from the past three years. After that, even "write a parent circular about next Friday's outing being postponed" came back in house format.

Note: the target throughout was format and tone, never facts. Date, venue and fee still had to come from the teacher's prompt or the knowledge base. Hardware threshold in the LoRA and QLoRA lesson.

A decision order

When output does not meet requirements, work through this order:

1. Is the prompt too vague? Add year level, length, format and prohibitions. Costs nothing, solves about sixty per cent.
2. Are school-specific facts missing? Build a knowledge base so it cites school documents. About thirty per cent, and reduces hallucination as a bonus.
3. Is tone or format persistently wrong despite a standardised prompt? Only now consider fine-tuning.
4. Do you want it to look things up and take actions itself? That is agents and tool calling, not fine-tuning.

Where people go wrong

• The mistake: fine-tuning to install school facts (marking criteria, school history, policy detail). Self-check: ask the fine-tuned model to "quote paragraph two on page three" — if it cannot cite reliably, use a knowledge base.
• The mistake: jumping to fine-tuning while every colleague writes a different prompt. Self-check: unify the template and trial for two weeks; the template itself is the best source of demonstration data if you fine-tune later.
• The mistake: checking only the target task after fine-tuning, with no regression test. Self-check: re-run five tasks the model previously handled well and confirm no catastrophic forgetting.

Try it once

Time: about 10 minutes. You need: any conversational AI tool.

1. Paste:

   For this message, do not answer my question at all. Simply imitate my sentence pattern and write three similar sentences.
   Please explain photosynthesis.
   

2. Watch: does the model comply? Does it "give in" and add an explanation after the three sentences?
3. Hand in: the full model reply plus one sentence from you: "Does this show instruction-following is innate from pre-training or taught afterwards? Why?"

You got it right if: most models comply first then add an explanation — instruction-following is a strong habit added by SFT, while the pre-training instinct to continue text is still there. Next: the side effect — it follows instructions, including the wrong person's — alignment with RLHF and DPO. Previous lesson: pre-training.