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

Short answer: pre-training teaches a model what language looks like. Fine-tuning teaches it what to do when a person asks it something, by showing it large numbers of human-written demonstration exchanges to imitate.

What an experienced teacher remembers changing

An ICT teacher of fifteen years remembers that, before roughly 2022, the language models he tried had an odd habit. Type "Please explain photosynthesis" and it would not explain anything. It would continue: "Please explain respiration. Please explain transpiration. Please explain…" It had read your sentence as an item on a worksheet and helpfully written more items.

That behaviour is entirely reasonable. As lesson 6 explained, the only task in pre-training is guessing the next word, and in text scraped from the web, the thing most likely to follow "Please explain photosynthesis" really is another question, not an answer.

So the model was not incapable of explaining. It did not know you wanted an explanation. Turning "does not know what you want" into "knows what you want" is what fine-tuning does.

The student teacher and the mentor

In classroom terms, a freshly pre-trained model is like an extremely well-read student teacher. He knows the syllabus, writes well and has read everything, but has never stood in front of a class. Tell him to "deal with this group" and he has no idea whether you mean take the register, start the lesson, or send someone outside.

How does a school train that person? Not by giving him more books, but by attaching him to an experienced mentor. The mentor demonstrates: this is what you say to a latecomer, this is how you reply to a complaining parent, when a student asks about homework you first ask which step they are stuck on. After a few hundred demonstrations, the student teacher has internalised what to do in each situation.

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

Concretely, the team assembles a large set of demonstration exchanges: an instruction paired with 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 then continues using the same mechanism as lesson 6, except that what gets covered up is the ideal response. Every wrong guess nudges the parameters, and after hundreds of thousands of rounds the model has acquired the habit of carrying out instructions rather than extending them.

Two things that get overlooked

First, the volume is small and the quality bar is brutal.

Pre-training runs on trillions of tokens. SFT typically uses tens to hundreds of thousands of demonstrations. Several orders of magnitude smaller, but 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 mentoring a student teacher — a mentor who cuts corners a few times will find the habit copied.

Second, fine-tuning changes behaviour, not knowledge.

This is where schools most often go wrong. Many panel heads' first instinct is to fine-tune on all the school's curriculum documents so the AI "knows our school".

It usually does not work. The model does see those documents during fine-tuning, but what it learns is what documents of that kind look like, not what page three of that particular document says. Quoting school facts accurately still calls for a school knowledge base, because a knowledge base can cite a source, can be updated, and takes effect the moment a document is revised — whereas changing a fine-tuned model means training again.

A summary worth taking into a meeting: fine-tuning teaches it how to say things; a knowledge base gives it what to say.

An example from the classroom

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

They did two things.

First, they exhausted the prompt route properly: the template moved into a shared folder, one colleague maintained it, and the must-include items became a checklist. That closed about eighty per cent of the gap at zero cost.

Second, for the remaining twenty per cent — mostly paragraph order and customary courtesies — they ran a light fine-tune on their on-premises model using two hundred real circulars from the past three years. After that, even a bare instruction like "write a parent circular about next Friday's outing being postponed" came back in house format.

Note that the fine-tuning target throughout was format and tone, never facts. The date, venue and fee of the outing still had to come from the teacher's prompt or from the knowledge base. The practical threshold and hardware for fine-tuning is covered in the LoRA and QLoRA lesson.

A decision order you can take into a meeting

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 of cases.
2. Are school-specific facts missing? Build a knowledge base so it cites school documents. Solves about thirty per cent, and reduces hallucination as a bonus.
3. Is the 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.

Try it yourself

To watch instruction-following behave like a taught habit, paste this into any AI tool:

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

Most models comply, but you will notice how much they want to answer — some write the three sentences and then cannot resist adding an explanation. That is a good illustration of instruction-following as an acquired and quite strong habit.

What this means for your classroom

1. Phrase requirements as instructions rather than descriptions. "Write a Primary 5 worksheet on…" works better than "I'm thinking about something for Primary 5", because the model was trained to carry out instructions.
2. Do not use fine-tuning to solve a facts problem. School facts belong in an updatable, citable knowledge base, or every curriculum revision means retraining.
3. Standardise prompts before discussing fine-tuning. A panel-maintained template costs nothing and is often sufficient — and it doubles as the best source of demonstration data if you later do fine-tune.
4. Run a regression test after fine-tuning. Re-test on tasks the model previously did well to confirm nothing else degraded.

Next

The model now follows instructions. That has a dangerous side effect: if someone instructs it to do something it should not, it will happily comply. Teaching it to keep to the rules is the next lesson.

Continue with alignment with RLHF and DPO. Previous lesson: pre-training, self-supervised learning and data cleaning.