---
title: "Fine-Tuning: definition and example – Carl-Uwe Würfel"
description: "Fine-tuning explained: how an AI model is trained further on your own examples, when it beats prompting and RAG, and where its limits are for a business."
url: https://carl-wuerfel.de/en/glossary/fine-tuning
author: Carl-Uwe Würfel
---

1. [Home](https://carl-wuerfel.de/en)
  2. [Knowledge](https://carl-wuerfel.de/en/knowledge)
  3. [Glossary](https://carl-wuerfel.de/en/glossary)
  4. Fine-Tuning

# Fine-Tuning

Fine-tuning is the additional training of a pretrained AI model on your own reviewed examples so that it handles a specific task with a set format, style or behavior.

## On this page

  1. How fine-tuning works at its core
  2. Full fine-tuning, LoRA and other methods
  3. What fine-tuning is good for, and what it isn't
  4. Prompting and source access first, fine-tuning second
  5. Risks: overfitting, forgetting, data protection
  6. What fine-tuning costs and what it locks you into
  7. Where the term stops working
  8. Frequently asked questions
  9. Sources

 _Figure: A pretrained model gets additional training on suitable, reviewed examples and is evaluated on separate data; one possible goal is a more consistent output format. Fine-tuning does not replace access to current sources, and whether it helps compared with prompts or other methods has to be measured._

**Example**

A company sorts incoming customer inquiries into one of a few fixed categories, such as billing, complaint, contract change or other. Inquiries from the archive that were already sorted correctly are reviewed, anonymized and split into a training set and an evaluation set. A smaller model is fine-tuned on them and goes live only once it sorts the evaluation set more reliably than the base model with a careful prompt. Cases where the model is unsure still go to a person.

A pretrained model has learned to understand and generate language from large amounts of text. Fine-tuning trains it further on a much smaller, purpose-built dataset. Each example pairs an input with the desired output, such as a customer inquiry and its correct category. The result is an adapted model that is deployed as its own version and answers more reliably in the desired form for exactly that task.

Fine-tuning is one of three ways to adapt a model to a task. Prompt engineering changes the instruction, [RAG](https://carl-wuerfel.de/en/glossary/rag) hands the model relevant documents with every request, and fine-tuning changes the model itself. Even the well-known AI assistants are products of fine-tuning: Anthropic describes Claude as already fine-tuned to be a helpful assistant. In a business context, the term usually means additional training on the company's own data.

Executives need the term because it is often misunderstood. “We'll train the model on our data” sounds like the way to teach an AI what the company knows. That is what fine-tuning is worst at. It shapes format, style and behavior; current knowledge should reach the model through access to sources. A decision about fine-tuning is also a decision about training data, upkeep and lock-in to one provider and one model version.

## How fine-tuning works at its core.

The process has five steps. It starts with choosing a pretrained model as the base. Then comes most of the work: collecting, reviewing and consistently formatting suitable examples. The examples are split into two sets. The first is used for training; the second stays untouched and is used only for evaluation.

In the third step, the model gets additional training on suitable, reviewed examples. Its parameters, the weights it has learned, are shifted step by step so that it becomes more likely to produce the desired output for the example inputs. The result is an adapted model that is called as its own version.

Finally comes evaluation on separate data. The adapted model is compared against the base model given a carefully built prompt. Only that comparison shows whether the training achieved anything.

Fine-tuning does not make a model fundamentally smarter. It learns patterns from the examples and repeats them. If the examples are inconsistent or wrong, it learns that too.

## Full fine-tuning, LoRA and other methods.

Full fine-tuning adjusts all of a model's parameters. That takes a lot of compute and memory and yields a full model copy per task. Parameter-efficient methods change only a small part. The best known is LoRA (low-rank adaptation), introduced in 2021 by Edward Hu and colleagues at Microsoft. LoRA freezes the original weights and trains only small additional matrices. The result is an adapter, a small add-on to the base model that, according to the paper, adds no inference latency. Several adapters for different tasks can share the same base model. Google describes the parameter-efficient option for its Gemini models as more resource-efficient and cost-effective; full fine-tuning can deliver better results on complex tasks but costs considerably more.

Methods also differ in the kind of examples they use. In supervised fine-tuning, each example shows the correct answer. In preference training, such as direct preference optimization (DPO), the model sees a better and a worse answer and learns to prefer the better one. In reinforcement fine-tuning, a grader scores the answers, for instance whether generated code runs without errors.

Ask vendors which method they plan to use and why. The answer shows whether the project has been thought through.

## What fine-tuning is good for, and what it isn't.

Fine-tuning is strong on narrow tasks that repeat often and whose correct result is easy to define. These include classification, meaning sorting texts into fixed categories, extracting details into a fixed schema, a more consistent output format and a particular tone. OpenAI and Google list similar [use cases](https://carl-wuerfel.de/en/glossary/use-case) in their guides.

A second reason is operational: a model that knows the task from training needs shorter instructions and fewer examples in the prompt. With very high volumes of similar requests, that can lower cost per call and response time. Sometimes a smaller fine-tuned model handles a narrow task as well as a large model with a long prompt.

Fine-tuning is weak on knowledge. Prices, rules, product data and contacts change, but a trained model only knows the state of its training data. It cannot cite the source of a statement, and it reproduces individual facts from a handful of examples unreliably. Tasks whose rules change often are a poor fit too, because every change means retraining.

A rule of thumb: fine-tuning for how an answer looks, access to sources for what it says.

## Prompting and source access first, fine-tuning second.

The providers recommend a clear order. OpenAI advises writing [evals](https://carl-wuerfel.de/en/glossary/evals) first, so performance can be measured, and notes that good [prompt engineering](https://carl-wuerfel.de/en/glossary/prompt-engineering) is often all you need. Google recommends starting with prompting when labeled data is limited or a prototype is needed quickly.

That suggests a simple path. First, build a fixed list of real tasks with known correct answers. Then write a careful prompt with clear instructions and a few examples in it (few-shot). If the task needs knowledge from documents, add source access through RAG. Only if that setup measurably falls short is fine-tuning the next step.

The reason is effort. A prompt can be changed in minutes and takes effect immediately. Training needs prepared data, an evaluation and a new model version, and every change to the task starts the cycle over. Whether fine-tuning helps compared with prompts or other methods therefore has to be measured, on the same task list and against the same standard.

## Risks: overfitting, forgetting, data protection.

[Data quality](https://carl-wuerfel.de/en/glossary/datenqualitaet) decides the outcome; Google states that well-labeled data matters more than volume. The model learns the errors, contradictions and outliers in the examples along with everything else.

With too few or too one-sided examples, the model memorizes the training data instead of generalizing the task. This is called overfitting, and it shows when the model shines on the training examples but does worse on new cases. That is exactly what the separate evaluation set is for. A second risk is that the model loses abilities it had before; researchers call this catastrophic forgetting. Evaluation should therefore also look at tasks outside the training goal.

An adapted model can still hallucinate. Fine-tuning does not replace reviewing the output.

If the examples contain personal data, training needs a legal basis under data protection law such as the GDPR. A model can reproduce parts of its training data, and individual records cannot be removed from a trained model in a targeted way; if in doubt, it has to be retrained. It is safer to anonymize examples before training and leave confidential content out.

## What fine-tuning costs and what it locks you into.

Costs fall into four buckets. The first, and often the largest, is data work: collecting, reviewing, standardizing and splitting examples. The second is the training itself, which grows with model size, data volume and method. The third is running the adapted model, whether with a provider or on your own infrastructure. The fourth is upkeep: when a better base model comes out or the task changes, the model is retrained and re-evaluated.

Then there is lock-in. A fine-tuned model is tied to one provider and one model version. When the base model is retired, the adapted model goes with it. The market is shifting as well: as of November 2026, Anthropic's Claude API offers no fine-tuning, and OpenAI is winding down its fine-tuning platform. New users no longer get access, and existing fine-tuned models remain available until their base model is deprecated. Google offers tuning for Gemini models on its cloud platform (Gemini Enterprise Agent Platform since April 2026, formerly Vertex AI). Open-weight models can be trained on your own or rented infrastructure.

The training and evaluation data are therefore the real asset of a fine-tuning project. Keep them, and you can retrain on another model and compare.

## Where the term stops working

Fine-tuning does not replace access to current sources: an adapted model only knows the state of its training data, cites no sources and can still hallucinate. Whether it helps compared with prompts or other methods has to be measured. It is distinct from prompt engineering, which only changes the instruction, from RAG, which pulls knowledge from documents on every request, and from pretraining, which creates the model in the first place.

## Frequently asked questions

What is the difference between fine-tuning and RAG?

Fine-tuning changes the model itself through additional training, while RAG hands it relevant documents with every request. Fine-tuning shapes format, style and behavior; RAG supplies current knowledge with a cited source. The two can be combined.

Can fine-tuning teach an AI model company knowledge?

Only unreliably. A model mainly learns patterns from examples and does not reliably reproduce individual facts, and once the knowledge changes, the model is out of date. For company knowledge, source access through RAG is usually the better route.

How many examples does fine-tuning need?

There is no fixed minimum; it depends on the model, the method and the task. For tuning its Gemini models, Google cites a dataset of around 100 examples or more as a guideline and stresses that quality matters more than quantity. A separate set for evaluation comes on top.

When is fine-tuning worth it?

When a narrow task comes up very often, its correct result is easy to define and a good prompt with source access measurably falls short. Another reason is cost and response time at high volume, when a smaller adapted model solves the task as well as a large one. Reviewed examples and a separate evaluation are prerequisites.

Can you fine-tune ChatGPT or Claude?

As of November 2026, hardly. According to Anthropic, the Claude API offers no fine-tuning, and OpenAI is winding down its fine-tuning platform and no longer admits new users. Google offers tuning for Gemini on its cloud platform, and open-weight models can be trained in-house.

What happens to a fine-tuned model when a new base model comes out?

It stays where it was and does not benefit from the improvements. To get them, you have to retrain and re-evaluate on the new model. When the old base model is retired, the model trained on it usually goes with it.

Can you fine-tune on customer data?

If the examples contain personal data, training needs a legal basis under data protection law such as the GDPR. Because a model can reproduce training data and individual records cannot be removed in a targeted way, anonymizing before training is the safer route. Depending on the use case, obligations under the EU [AI Act](https://carl-wuerfel.de/en/ai-act) may apply as well.

## Further reading

  * [**AI glossary**](https://carl-wuerfel.de/en/glossary)
  * [**AI automation**](https://carl-wuerfel.de/en/ai-automation)
  * [**LLM: Large Language Model**](https://carl-wuerfel.de/en/glossary/llm)
  * [**Prompt Engineering**](https://carl-wuerfel.de/en/glossary/prompt-engineering)
  * [**RAG: Retrieval-Augmented Generation**](https://carl-wuerfel.de/en/glossary/rag)

## Sources

  1. [Anthropic: Glossary (documentation)](https://platform.claude.com/docs/en/about-claude/glossary)
  2. [Hu et al.: LoRA: Low-Rank Adaptation of Large Language Models, 2021](https://arxiv.org/abs/2106.09685)
  3. [Google Cloud: Introduction to tuning (documentation)](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/tuning)
  4. [Rafailov et al.: Direct Preference Optimization: Your Language Model is Secretly a Reward Model, 2023](https://arxiv.org/abs/2305.18290)
  5. [OpenAI: Model optimization (documentation)](https://developers.openai.com/api/docs/guides/model-optimization)
  6. [European Union: Regulation (EU) 2016/679 (General Data Protection Regulation), 2016](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679)
  7. [European Union: Regulation (EU) 2024/1689 (Artificial Intelligence Act), 2024](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)