← back to the archiveCover illustration for “Fine-tuning: adapting a trained model with your examples”
TERM8mo ago·Published · Updated ·by Andy Padia

Fine-tuning: adapting a trained model with your examples

In short: Fine-tuning continues training a pre-trained model on your own examples to change its behaviour, style or task skill, and it suits behaviour more than facts.

Fine-tuning is further training of an already pre-trained model on a smaller, task-specific dataset, so its weights change to favour the behaviour you want: a format, a tone, domain vocabulary or a narrow skill such as classification. Prompting changes only the input, and RAG adds knowledge at query time. OpenAI's supervised fine-tuning guide, read in November 2026, accepts as few as 10 examples.

Fine-tuning in plain words

Prompting is giving a skilled cook a recipe card each time. Fine-tuning is sending them on a short course so they cook your house style without the card.

How it works

  1. Prepare examples: inputs paired with ideal outputs. OpenAI's guide recommends starting with 50 well-crafted demonstrations and then evaluating the results.
  2. Train starting from the base model's weights.
  3. Evaluate on data you held back, against the base model plus a good prompt. The same guide says to invest in fine-tuning only after setting up evals.
  4. Deploy the tuned model or adapter, and re-evaluate when the base model changes.

With open models you run this yourself. Hugging Face's TRL library provides a supervised fine-tuning trainer and lists PEFT, the home of adapter methods such as LoRA, among its integrations.

Example

As an illustration, an insurer wants claim notes turned into a strict internal summary format with its own codes. A good prompt gets most outputs right but keeps slipping on the codes. Fine-tuning a small model on a few thousand approved examples is the next thing to test, and the shorter prompt would also cut cost per call.

Where you'll meet it

  • Fine-tuning APIs from model providers
  • Open-source tooling such as Hugging Face TRL and Unsloth
  • Domain models in legal, medical and finance
  • Distilling large-model behaviour into small models

What it changes for you

For teams, fine-tuning is a data project first: collecting, cleaning and approving examples is most of the work. For architects, a tuned model is an artefact to version, evaluate and retrain when the base model is retired. For most knowledge problems, RAG is the better tool; fine-tune for behaviour, not facts.

Fine-tuning vs RAG

RAG supplies facts at question time and updates instantly when documents change. Fine-tuning changes behaviour and needs retraining to change.

Common mistake

My view: fine-tuning to teach a model company facts. It memorises poorly, cannot cite sources and goes stale. Put facts in retrieval and use fine-tuning for format, tone and task skill.

LoRA · Retrieval-augmented generation · Distillation · Golden dataset · Few-shot prompting

From the archive: Fine-tune to change behaviour, use RAG to change knowledge · Choose prompts, RAG, tools or fine-tuning

Put facts in retrieval, and fine-tune only for format, tone and task skill.

#field-guide#fine-tuning#models#training

Sources

  1. developers.openai.com/api/docs/guides/supervised-fine-tuning
  2. huggingface.co/docs/trl/index
← older drop
Human-in-the-loop: where a person approves AI actions
newer drop →
Regression evals: catching AI changes that break things

related drops

explore all 557 drops →
← back to the archiveday 175