STAGE 0 · 4.3 · 3 NODES

From base model to assistant

A freshly trained model can't answer a question — it can only continue text. Here's the small, cheap step that turns that raw "base" model into the helpful assistant you actually talk to.

THE 3 STEPS

01Two behaviours 02The training pairs 03What actually changes

NODE 01 / 03

Same prompt, two very different replies

A base model and an assistant can share the exact same brain — but ask them the same thing and you get two different behaviours.

A base model just continues text. It was trained to predict the next word, so it treats your prompt as the start of a document and keeps writing.
An assistant answers. Fine-tuning taught it that a prompt is a request, so it responds directly and then stops.
Same knowledge, different reflex. Flip base ↔ assistant on each prompt and watch the reply style change — the facts don't.

BASE vs ASSISTANT · SAME PROMPT

Pick a prompt, then flip base ↔ assistant — the reply style changes, the facts don't.

Base model — trained only to predict the next word. It continues your text like a document, often drifting into lists, repeats, or brand-new questions.

NODE 02 / 03

The fuel: instruction → response pairs

Turning a base model into an assistant takes examples — hundreds of thousands to millions of them — each pairing a request with a good reply.

Every example is a pair: an instruction (the prompt) and the response you want the model to give.
The model learns the mapping, not the specific facts — when asked, produce a helpful reply in this shape.
Pairs teach behaviour: answer the question, follow the format, and refuse what's harmful.

ONE TRAINING PAIR

Scroll or tap through the kinds of pairs an assistant learns from.

Answer a question — the most common pair. Teaches the model to respond directly with correct information.

NODE 03 / 03

What fine-tuning changes — and what it doesn't

Fine-tuning is a light touch. It reshapes how the model behaves without rewriting what it knows.

Knowledge barely moves. Facts, grammar and reasoning were learned in pretraining — fine-tuning leaves them almost untouched.
Behaviour jumps. The model learns to follow instructions, use a helpful format, and refuse harmful asks.
Cheap and fast. Instruction tuning uses a tiny fraction of the pretraining data and compute — it's a polish, not a rebuild.

WHAT MOVES · BASE → ASSISTANT

Flip base ↔ assistant — knowledge stays put, behaviour is what moves.

Base — strong knowledge, but it continues text instead of helping. The raw pretrained model.

DELTA
KNOWLEDGE90 / 100
BEHAVIOUR14 / 100
Knowledge stays flat; helpful behaviour is what fine-tuning adds.

EXPLAIN IT BACK

You fine-tune a base model on instruction → response pairs and it becomes a helpful assistant. Did it get smarter?

NEXT: WHY ALIGNMENT

Instruction tuning makes a model helpful. But helpful isn't the same as safe or honest — a model can follow a bad instruction just as eagerly as a good one. Next, why alignment: teaching a model to follow the right instructions, not just any instruction.

← How models learn Why alignment →

Language: English