What is fine-tuning?
Fine-tuning is further training of an already trained AI model on a specific, smaller dataset so that the model is better tailored to a particular use case.
DEFINITION
In fine-tuning an AI model, an already trained base model serves as the foundation and receives further training on data from a specific application context. A general language model learns the language, style, terminology and task logic of the respective domain. For example: a general LLM knows medical terms from its training dataset. A model fine-tuned on clinical reports answers more precisely, in the right style and with relevant knowledge. Fine-tuning is more resource-intensive than prompt engineering, but less demanding than training a model from scratch. The decision for fine-tuning makes sense when consistently specific output behaviour is needed that cannot be reliably achieved with prompts alone. An alternative is Retrieval-Augmented Generation (RAG), which does not require adjusting model weights.
CONNECTIONS
Leadership
Fine-tuning works like institutionalised feedback: the model learns from rated examples just as organisations learn from leadership feedback. Anyone who uses fine-tuning must build a clear feedback culture, because the quality of training data directly determines model quality.
Agility
Fine-tuning projects need clear acceptance criteria: when is a fine-tuned model good enough for deployment? A Definition of Done for AI outputs helps make quality measurable and avoid rework.
Project Management
Fine-tuning projects have clear phases: data curation, training, evaluation, deployment. Those who draw lessons learned from each cycle improve not only the model but also the next training process.
KEY POINTS
- Fine-tuning adapts a pre-trained model to a specific use case.
- It permanently changes model weights, unlike RAG.
- Less demanding than training from scratch, but more expensive than prompting.
- Suitable for consistently specific output behaviour.
- The quality of the fine-tuning dataset is decisive for result quality.
EXAMPLE
A customer service team at a telecommunications company has thousands of resolved support cases with corresponding answers. They fine-tune a language model on this data. The result: the model responds in the company tone, knows the company’s own products well and formulates precise answers to typical customer problems without a detailed prompt being needed each time.
MISCONCEPTIONS
Is fine-tuning the same as prompt engineering?
No. Prompt engineering changes nothing in the model, only in the request. Fine-tuning permanently changes the model’s weights through further training.
Is fine-tuning necessary to use an LLM in a company?
No. For many use cases, good prompts or RAG are sufficient. Fine-tuning only pays off when consistently specific behaviour is needed that other methods cannot deliver reliably.