# Fine-Tuning

> Fine-tuning is the process of further training an existing language model on a set of example inputs and outputs so it adopts a particular style, format or task behavior. It changes the model's weights, producing a customized version. Fine-tuning teaches patterns well but is a poor way to add facts that change often, which is usually better handled with retrieval.

## Why it matters for your business

Businesses often assume they need to train a model on their data, when retrieval and good prompts would solve the problem faster and more cheaply. Fine-tuning requires a clean, consistent set of examples, often starting at 50 to 100 and growing into the hundreds for harder tasks, and usually has to be redone to move to a newer base model. It makes sense for narrow, high-volume tasks where consistent format or tone matters, or where a smaller fine-tuned model could replace a larger, more expensive one. Try prompting and retrieval first, and measure with evals before deciding.

## Example

An ecommerce brand generates thousands of product descriptions a month. Prompting a general model produced acceptable copy, but the tone drifted from the brand voice. The team fine-tunes a smaller model on several hundred descriptions written by its copywriters, then compares it against the prompted larger model on a review set before switching production traffic.

Related terms: https://www.lumeor.studio/glossary/rag-retrieval-augmented-generation/, https://www.lumeor.studio/glossary/prompt-engineering/, https://www.lumeor.studio/glossary/evals-ai-evaluation/, https://www.lumeor.studio/glossary/llm-large-language-model/, https://www.lumeor.studio/glossary/inference/

---

Source: https://www.lumeor.studio/glossary/fine-tuning/