What is fine-tuning?

Training vs. fine-tuning

Training a large language model costs tens of millions of euros in computing power. This is not feasible for most organizations. Fine-tuning offers an alternative: you take an already-trained base model and adapt it for your specific application with a relatively small dataset.

The base model has already built up broad knowledge through pre-training. Fine-tuning adds a layer of specialization on top of that, without having to relearn everything.

When is fine-tuning useful?

• You want to train a model on jargon or style from your sector (legal, medical, technical)
• You want consistent behavior: always the same format, the same tone
• You want to process sensitive data without using an external API
• You want better performance on a specific task than a generalist model provides

Illustration created with Canva AI

How does fine-tuning work?

1. Compile training data — examples of input-output pairs demonstrating the desired behavior
2. Choose a base model — an open-source model (Llama, Mistral) or via an API (OpenAI, Anthropic)
3. Train — the model is adjusted on your dataset via gradient descent
4. Evaluate — testing on a separate dataset to detect overfitting
5. Deploy — putting the fine-tuned model into production

Common mistakes

• Too little data — a few hundred examples is often insufficient; thousands are better
• Poor data quality — inconsistent or erroneous examples degrade the model
• Overfitting — the model memorizes the training data instead of generalizing

Fine-tuning vs. RAG

Fine-tuning is not always the best choice. If you want the model to have access to current or changing information, RAG (Retrieval-Augmented Generation) is often better. Fine-tuning is suitable for style and behavior; RAG is suitable for knowledge and facts.