What is fine-tuning an LLM, and when should you do it?

Fine-tuning teaches a model how to behave; it is a poor way to teach it what to know. Most teams that reach for fine-tuning first would get a better result, sooner, from clearer prompts, examples or retrieval.

How does fine-tuning work?

Fine-tuning takes a model that has already been pretrained and runs more training on a much smaller dataset, usually hundreds to tens of thousands of examples of inputs paired with the outputs you want. The training nudges the model's weights so that, given similar inputs in future, it produces outputs like your examples. This is called supervised fine-tuning (SFT). Chat models are themselves built this way: a base model is fine-tuned on instruction-and-response examples, then refined further with preference training such as RLHF.

What is LoRA?

LoRA (low-rank adaptation), published by Microsoft researchers in 2021, is a way to fine-tune without updating all of a model's weights. It freezes the original model and trains small extra matrices added to certain layers. The paper reported cutting the number of trainable parameters for GPT-3 by about 10,000 times, with quality comparable to full fine-tuning on the tasks tested. QLoRA, from 2023, combined LoRA with a compressed 4-bit copy of the base model, making it possible to fine-tune a 65 billion parameter model on a single large GPU.

Full fine-tuning

Updating every weight in the model; most flexible and most expensive

PEFT

Parameter-efficient fine-tuning, a family of methods that train only a small number of extra parameters

LoRA

A popular PEFT method that trains small low-rank matrices alongside frozen weights

Adapter

The small set of trained extra weights, which can be swapped in and out of one base model

Overfitting

When a model memorises its training examples and gets worse on new inputs

Should I use fine-tuning, RAG or prompting?

Use prompting first, RAG when the model needs facts it does not have, and fine-tuning when you need consistent behaviour that prompts cannot reliably produce.

Fine-tuning is weak at adding knowledge. A model tuned on your product manual may pick up its style and some facts, but it will still mix them with things it made up, and you cannot easily update it when the manual changes. Retrieval handles facts better because the source is visible and replaceable.

What does fine-tuning need?

1. 1
   A clear goal

   Write down the behaviour you want and how you will measure it before collecting data.
2. 2
   A baseline

   Measure the best prompt-only version on your eval set, so you know whether tuning helped.
3. 3
   Good examples

   Hundreds of high-quality, varied examples usually beat thousands of noisy ones. Every flaw in the data is learned too.
4. 4
   A held-out test set

   Keep examples out of training to check the model generalises instead of memorising.
5. 5
   A plan for updates

   Base models improve every few months; decide how you will re-tune or re-evaluate when they do.

Hosted APIs from OpenAI, Google and others let you fine-tune some of their models by uploading a file of examples. Open-weight models such as Llama, Mistral and Qwen can be tuned on your own hardware with LoRA.

• ✓Do: Fine-tune to make outputs consistent, shorter or cheaper to produce
• ✓Do: Compare against a well-prompted base model on the same test set
• ✓Do: Keep the training data under version control with the resulting model
• ✕Avoid: Fine-tune to teach facts that change
• ✕Avoid: Skip the baseline; many fine-tunes turn out no better than a good prompt
• ✕Avoid: Tune on examples you have not read