Supervised and Instruction Fine-Tuning

Supervised fine-tuning (SFT) means: show the model many examples of “given this input, produce that output,” and update weights so it copies the desired pattern. Instruction fine-tuning (IFT) is the chat-friendly version of that idea — teach the model to follow natural-language instructions.

Intuition

Simple analogy: SFT is like tutoring with answer keys. Each row says what good looks like. After enough examples, the model internalizes the habit — not only for that exact wording, but for similar asks.

How it works

Supervised fine-tuning (SFT)

You need clear inputs and target outputs.
Format must be consistent (same schema, same role tags if you use chat).
Loss usually focuses on predicting the target tokens (the answer), not forcing the model to memorize the user’s question as if it wrote it.

Instruction fine-tuning (IFT)

Instruction data often looks like:

A user instruction (and maybe context)
An ideal assistant response

Training teaches the model: when someone asks in natural language, answer helpfully in the expected style.

One training row, written out in full:

{
  "messages": [
    {"role": "system",    "content": "You are a support agent for a bank."},
    {"role": "user",      "content": "I lost my debit card. What should I do?"},
    {"role": "assistant", "content": "I'm sorry to hear that. Please block the card immediately from the app under Cards > Block, then request a replacement. Your new card arrives in 3-5 working days."}
  ]
}

Two details matter here:

The same role structure must be used during training and when the model is later served. If training used system/user/assistant and production sends a plain string, the model sees an unfamiliar shape and quality drops.
The loss is calculated on the assistant text only. The model is being taught to produce that reply, not to invent the customer's question.

SFT vs IFT (practical view)

	SFT (broad)	IFT (instruction-focused)
Data	Any labeled input→output pairs	Instruction / chat-style pairs
Goal	Task skill (classify, extract, draft…)	Follow instructions in assistant form
Overlap	IFT is a common modern form of SFT for chat models	Same training idea; different data shape

Many people say “SFT” when they mean instruction tuning a chat model. That is fine — just know which data format you are using.

Effect of instruction fine-tuning

After IFT, models usually become better at:

Following directions
Producing useful assistant-style answers
Sticking to requested formats

They still need good data. Garbage instructions teach garbage habits.

How much data do you need?

There is no magic number, but a useful starting frame:

Goal	Rough scale of examples
Lock in one output format	Hundreds
Teach a task with real variety	A few thousand
Broad instruction-following from a base model	Tens of thousands and up

Quality beats quantity almost every time. 500 carefully checked examples usually beat 5,000 scraped ones, because the model copies whatever it is shown — including the mistakes.

What goes wrong

Mixing random formats in one JSONL file so the model never sees a stable pattern.
Training on answers only, with unclear instructions.
Expecting IFT to install fresh facts that belong in RAG.
Letting a few sloppy examples slip in. If 5% of your answers are rude or truncated, the model learns that rudeness and truncation are sometimes correct.
Using one chat template during training and a different one in production, then wondering why the model got worse after deployment.

One-line summary

Supervised fine-tuning teaches from labeled examples; instruction fine-tuning is that idea applied to chat-style “follow my request” data.

Key terms

SFT — Supervised fine-tuning on labeled pairs.
Instruction fine-tuning (IFT) — SFT focused on instruction→response behavior.
Chat template — The role/format wrapper used at train and serve time.