Contents
At a glance
- Category
- Agent concept - model adaptation
- Teaches
- Format, tone, structure, a narrow skill or classification boundary
- Does not teach
- Facts - use retrieval for knowledge that must be current and citable
- Usual method
- Parameter-efficient tuning such as LoRA rather than full retraining
- Dataset
- A few hundred consistent, checked examples beats thousands of noisy ones
- Hidden cost
- Ongoing maintenance - re-tuning on base model deprecation and format changes
What fine-tuning is
Fine-tuning takes a pre-trained model and continues training it on a set of your own input-output examples, adjusting its weights so the behaviour you demonstrated becomes its default. Modern practice usually means parameter-efficient methods such as LoRA, which train a small set of additional weights rather than the whole model, cutting cost and time by orders of magnitude.
The single most important thing to understand is what it teaches. Fine-tuning teaches how to respond - format, tone, structure, a specialised classification boundary. It is a poor way to teach what is true. Facts baked into weights cannot be cited, cannot be updated without retraining, and blend unpredictably with what the model already believed. For knowledge, use RAG.
When it is the right tool
Four situations justify it. You need output in a rigid format that prompting keeps almost getting right. You need a house tone that takes three hundred words of prompt to describe and still drifts. You have a classification task with a subtle boundary that examples convey better than rules. Or you want to run a much smaller, cheaper model at the quality of a larger one on one narrow task - which is the case with the clearest financial return, and is closely related to distillation.
Before any of them, exhaust the cheaper options: a better system prompt, worked examples in the prompt, structured output mode, and retrieval. Most fine-tuning projects that get abandoned were solving a prompt problem the expensive way.
Cost, data and effort
The training bill is rarely the issue - with parameter-efficient methods, a few hundred examples on a small model is inexpensive on hosted platforms and cheap on your own hardware for open-weight models. The cost is the dataset. You need consistent, correct examples, and quality dominates quantity: a few hundred carefully checked pairs beat thousands of scraped ones, because the model will faithfully learn any inconsistency you leave in.
The other cost is permanence. A fine-tuned model is a version you now own and must maintain: re-tune when the base model is deprecated, re-tune when your format changes, and keep the eval set that proves it still works. That maintenance is the real reason to be sure prompting cannot do the job.
What you should know
Your best training data is a by-product of running the process manually. Every corrected draft from a human-in-the-loop step is a labelled example of input and desired output. Capture those from day one and the dataset assembles itself; try to write one from scratch later and it becomes a project.
Check the data terms too. Fine-tuning means uploading your examples to a provider, so confirm what happens to them, whether they are used for training, and where they are stored - see zero data retention and DPA. With open-weight models you can fine-tune entirely on infrastructure you control, which is often the deciding factor for regulated work.
The alternative to check first
Before fine-tuning, price the cheap-model option, because the gap between tiers has widened enough to change the answer. In August 2026 gpt-5.6-luna was published at $0.20 per million input tokens and $1.20 output against gpt-5.6-sol at $5.00 and $30.00 - one twenty-fifth of the price, with the same 1,050,000-token context window.
A great deal of what fine-tuning was reached for historically - matching a house format, sticking to a fixed label set, producing consistent structured output - is now achievable by pointing a cheap model at a clear schema and a good prompt. That path has no training cost, no dataset to maintain and no retraining when the base model is superseded, which is the hidden recurring cost of a fine-tune.
Fine-tuning still wins where the task depends on knowledge or style that cannot be expressed in a prompt of reasonable length, or where latency and per-call cost at extreme volume justify the maintenance. Decide it with an eval set that compares both, not by reputation.
Fine-tuning FAQ
What does fine-tuning actually do?
It continues training an existing model on your input-output examples so the demonstrated behaviour becomes its default. In practice that means locking in a format, a tone or a narrow classification skill without having to describe it at length in every prompt.
Should I fine-tune or use RAG?
RAG for knowledge, fine-tuning for behaviour. If the problem is that the model does not know your prices, policies or products, retrieval is the answer and it stays current. If the problem is that it will not reliably produce your format or tone, fine-tuning is the answer.
How many examples do I need to fine-tune?
Fewer than people expect, and quality matters far more than volume. A few hundred consistent, correct pairs typically outperform thousands of inconsistent ones, because the model learns your inconsistencies just as faithfully as your intentions.
How much does fine-tuning cost?
With parameter-efficient methods, the training run itself is usually modest on hosted platforms and cheap on your own hardware for open-weight models. The real costs are assembling a clean dataset and maintaining the tuned model when the base model changes underneath it.
Can fine-tuning make the model know my products?
Badly. Facts learned in weights cannot be cited, go stale immediately, and mix unpredictably with pre-existing knowledge. Put the product data in a document, retrieve it, and let the model read it - that stays accurate the moment you edit the source.
Where do I get fine-tuning data without a big labelling project?
From your human review step. Every draft a person edited or rejected is a pair of what the model produced and what was actually wanted. Capture both sides from the start and the dataset builds itself as a by-product of normal operations.
Primary source
OpenAI - model optimization and fine-tuning
Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.
Want this built rather than explained?
Knowing what fine-tuning means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.
