Large AI models are trained on general data with enormous effort. Such a foundation model handles many tasks reasonably well but is not tailored to the specifics of a particular use case. This is where fine-tuning comes in: instead of training a model from scratch, an existing one is further adapted with comparatively few suitable examples.
For companies, the main question is when this effort is worthwhile and when simpler approaches are sufficient.
How does fine-tuning work?
The basis is a dataset of examples that show how the model should respond. For language models, these are usually pairs of input and desired output, such as a customer inquiry and the matching category, or a bullet-point note and the report written from it. Training is continued with these examples, and the model's parameters shift toward the desired behavior.
Two variants are common:
- Full fine-tuning: All parameters are adjusted. This is compute-intensive and costly for large models.
- Parameter-efficient fine-tuning: Only a small set of additional parameters is trained, for example with the LoRA method. This saves computing power and memory.
Some AI model providers offer fine-tuning as a cloud service. Open models can also be adapted on your own hardware.
What fine-tuning is used for
- Consistent output formats, such as structured data for further processing
- A fixed writing style or a specific tone
- Classification tasks with company-specific categories
- Making smaller, cheaper models more capable for a narrowly defined task
What companies should keep in mind
- Simpler approaches first: Good prompt engineering or supplied documents often achieve the goal faster and more cheaply.
- Data quality: Incorrect or contradictory examples are learned as well. A few carefully checked examples are often more valuable than large, unchecked volumes.
- Personal data: Whatever goes into the training data can hardly be removed from a model in a targeted way. Such data should be removed or pseudonymized beforehand.
- Evaluate measurably: Check against the same test cases before and after training whether the result actually improves.
- Plan for maintenance: When a new base model is released, the fine-tuning usually has to be repeated.
How it differs
Fine-tuning changes a model's behavior but is only of limited use for teaching it current factual knowledge. Retrieval-augmented generation is usually better suited for this, because documents can be swapped at any time.
View AI consulting