Fine-tuning: retraining an AI model on your own examples

The first question about fine-tuning is not how, but whether. Fine-tuning means taking a model that has already been trained and training it further on your own examples, so that it performs a specific task, format or tone more reliably. Google’s glossary calls it a second, task-specific training pass on a pre-trained model.

Three ways to change what a model does

Approach	What changes	Suited to
Prompting with examples	Nothing in the model, only the instructions.	trying out a task, small volumes
Retrieval-augmented generation, based on embeddings	Nothing in the model; relevant documents are supplied with each request.	knowledge that changes, answers from your own documents
Fine-tuning	The model’s weights.	one job repeated at high volume

OpenAI names three advantages of fine-tuning over prompting alone: you can provide more examples than fit into a single request, shorter prompts reduce cost and latency at scale, and you can train on proprietary data without sending it along with every request.

The main methods

• Supervised fine-tuning: examples of prompts paired with the correct response.
• Direct Preference Optimization (DPO): for each prompt, a better and a worse response.
• Reinforcement fine-tuning: the model answers, the answers are graded, and higher-scoring reasoning is reinforced.
• Vision fine-tuning: the same idea applied to image inputs.

These are the methods OpenAI documents for its own models. For open models, LoRA, short for low-rank adaptation, has become a common way to keep costs down: it freezes the original weights and trains small additional matrices inserted into the layers, an approach published by Hu and colleagues in 2021.

A scenario that goes wrong

Suppose a company fine-tunes a model on three years of its own support replies, so that answers sound like the team. Some of those replies quote prices that have since changed, and a few contradict each other. The tuned model repeats both, fluently, because nothing in the training told it which replies were wrong.

Nobody notices at first, because nobody held back a set of test questions to compare the tuned model with the original. A year later a new generation of the base model arrives, and the tuning has to be repeated on it. The replies also contained customer names, so the training data needed a legal basis under data protection law before any of this began. And if the company later offers the tuned model to others, the company’s role under the EU AI Act may change, which is worth checking before release.

A cheap first step

Before deciding anything, write down twenty real inputs and the outputs you expect for each. That list is the test set for prompting, retrieval and fine-tuning alike, and it often shows that a well-built prompt already gets most of the way. Working out which approach fits, including whether a local model makes sense, is part of an AI implementation.

Related terms

Term	What it means
Base model	The pre-trained model that fine-tuning starts from.
Evaluation set	Examples held back from training to judge whether a change actually helped.
LoRA and QLoRA	Resource-efficient fine-tuning methods.

Sources

• Google for Developers, Machine Learning Glossary, entry “fine-tuning”, accessed 14 September 2026
• OpenAI, Model optimization, accessed 14 September 2026
• Hu et al., “LoRA: Low-Rank Adaptation of Large Language Models”, arXiv, 2021

This page expands an entry from the minoka AI glossary, which covers many more terms in brief.