Fine-tuning adjusts a model's internal weights, unlike prompting, which only changes the input. It is good for teaching consistent behaviour and tone. It is less reliable for adding new facts, where retrieval methods usually work better.
How it works
A foundation model has already learned general language or image skills from huge datasets. Fine-tuning continues training on your own examples, usually pairs of inputs and ideal outputs, such as a customer question and the perfect reply. The model's predictions are compared with your target answers and its weights are nudged to reduce the gap, using the same gradient descent process as original training but with far less data and compute. Quality matters more than quantity: a few hundred to a few thousand clean, consistent examples is a common starting range. Full fine-tuning updates every parameter, which is costly for big models. Parameter-efficient methods such as LoRA freeze the original weights and train small add-on matrices instead, cutting memory needs dramatically. Many providers offer fine-tuning through an API: you upload a file of examples, they train a private version and you call it like any other model. Related methods include instruction tuning, which teaches models to follow requests, and RLHF, which uses human preference ratings.
Why it matters
Prompts can only go so far. If you need hundreds of product descriptions in exactly your brand voice, a strict output format every time or a model that handles specialist jargon, fine-tuning can deliver more consistent results with shorter prompts, which can lower cost per request. It can also let a smaller, cheaper model match a larger one on a narrow task. For most teams, though, it is worth trying good prompts and RAG first.
Where you will see it
- Teaching a model your brand's tone for product descriptions or social captions.
- Making a model always return a specific JSON or report format.
- Classifying support tickets or comments into your own categories.
- Adapting an image model to draw a consistent character or style.
Limits and common misconceptions
- It is not a reliable way to add facts; models may still hallucinate, and knowledge goes out of date.
- Poor or inconsistent training examples teach poor behaviour, and fine-tuning can weaken some general skills.
- You must retrain when the base model changes, and training data may contain sensitive information.
How it differs from related terms
- Fine-tuning vs RAG (retrieval-augmented generation) RAG gives a model fresh information at question time without changing it, while fine-tuning changes the model's weights to alter how it behaves.
- Fine-tuning vs Prompt engineering Prompt engineering shapes output by changing instructions each time, whereas fine-tuning bakes the behaviour into the model through extra training.
Frequently asked questions
Should I fine-tune or use RAG?
Use RAG when the model needs specific, changing information such as documents, prices or policies. Use fine-tuning when you need consistent style, format or behaviour that prompts cannot deliver. Many production systems combine both: a fine-tuned model for tone, plus retrieval for facts.
How much data do I need to fine-tune a model?
Less than many people think. Providers often suggest starting with dozens to hundreds of high-quality examples and adding more if results improve. A few hundred to a few thousand consistent examples is common for business tasks. Clean, varied, representative examples matter more than raw volume.
What is LoRA?
LoRA stands for low-rank adaptation. Instead of updating all of a model's weights, it trains small extra matrices that sit alongside the frozen original. This uses far less memory and storage, and the resulting adapters are small files that can be swapped in and out.
Related terms
- Dubbing (AI) Using AI to translate a video's speech into another language and re-voice it, often in a clone of the original speaker's voice.
- Embedding A list of numbers (a vector) that captures the meaning of text, images or audio, so that similar items end up close together mathematically.
- Foundation model A large AI model pre-trained on broad data at scale that can be adapted to many different tasks, and that many apps are built on.
- Generative AI AI systems that create new content, such as text, images, audio, video or code, by learning patterns from large datasets and responding to prompts.
- GPU A graphics processing unit: a chip with thousands of cores running calculations in parallel, the main hardware for training and running AI models.