LLM Fundamentals and Why Fine-Tune

A language model is a model of text. Given the words so far, it estimates what comes next. A large language model (LLM) is that idea scaled up — many parameters, huge training data, usually built on transformers.

Intuition

Core job in one line: predict the next token.

Example: “The cat sat on the ___”

The model assigns probabilities to possible next words:

It then picks one, adds it to the sentence, and repeats the whole process for the following word. Everything an LLM does — chat, summarising, writing code — is this one step run over and over.

How it works

What “large” adds

Scale (parameters + data + compute) is why modern LLMs can:

• Write fluent text and dialogue
• Pick up a new task from a few examples in the prompt (in-context learning)
• Handle more reasoning, knowledge, code, and math than small models

Well-known families include GPT, Claude, Gemini, and Llama-style open-weight models.

Three common LLM shapes

An easy way to remember the difference:

• Encoder-only is a reader. It sees the whole sentence at once, like you reading a paragraph before answering a comprehension question. Perfect for “is this review positive or negative?”
• Decoder-only is a writer. It can only look backwards at what it has already written, like a person speaking a sentence out loud one word at a time. Perfect for chat.
• Encoder–decoder is a translator. One half reads the whole input, the other half writes a fresh output. Perfect for “turn this English paragraph into French.”

Almost every chat assistant you use today is decoder-only.

The usual training stages

Why fine-tuning is often necessary

A general pretrained model is a strong starting point, but many products need:

• Domain specialists (medicine, finance, legal drafting style)
• Stable formats and habits that prompting alone does not lock in
• Behavior that should stick across many future prompts, not only one clever instruction

Fine-tuning updates the model on task-specific data so those habits become part of the weights.

A concrete example

An insurance company wants every claim note summarised in exactly this shape:

CLAIM: <id>
LOSS TYPE: <fire | flood | theft | other>
AMOUNT: <number or unknown>
NEXT ACTION: <one sentence>

With prompting, the model gets it right most of the time, but on maybe 1 note in 12 it adds a friendly sentence at the top, or writes “Fire damage” instead of fire. At 4,000 notes a day, that is hundreds of broken records.

After fine-tuning on 2,000 correctly formatted examples, the format becomes a habit instead of a request. The prompt can then be short, because the model has already learned what the answer should look like.

That is the real test for fine-tuning: do you need this behaviour every single time, forever?

What goes wrong

• Treating a general chat model as already “done” for a specialist job.
• Fine-tuning to store facts that change every week (retrieval is usually better for that). Training last month’s price list into the weights means retraining when the prices move.
• Skipping the question “do we need weight updates at all?” Many teams spend a week fine-tuning a problem that a five-line prompt fix would have solved.
• Assuming the model “understands” instead of predicting. It has no idea whether the sentence it just produced is true — which is exactly why grounding and evaluation matter.

One-line summary

LLMs learn next-token prediction at huge scale; fine-tuning is how you turn that general skill into stable, task-specific behavior.

Key terms

• Language model (LM) — Model that predicts the next piece of text.
• LLM — A very large language model, usually transformer-based.
• Pretraining — Broad training before your task-specific work.
• Fine-tuning — Further training on task-specific data.