Fine-tuning
Definition
Fine-tuning further trains an existing model on company-specific examples so it adopts a style, format or classification behaviour. It changes how a model responds, not what current facts it knows, and is usually the last option after prompting and retrieval.
It pays off for narrow, high-volume, repetitive tasks with stable output formats — document classification, structured extraction, standardised correspondence.
It is a poor fit for knowledge that changes: prices, stock, contract terms. Those belong in retrieval or in a direct system query.
In practice
- Best for style, format and classification consistency
- Requires a curated, labelled set of company examples
- Adds a maintenance burden with every model generation
Related terms
- Large language model (LLM)A large language model is a statistical model trained on very large text corpora that predicts likely continuations of text. It can classify, summarise, translate and draft, and it is the reasoning component inside most AI agents — but it holds no live company data by itself.
- Retrieval-augmented generation (RAG)Retrieval-augmented generation is a method in which relevant company documents are searched first and the retrieved passages are given to the language model as context. The model answers from those passages, which makes answers current, source-linked and auditable.
- Prompt engineeringPrompt engineering is the practice of specifying, in writing, what a model must do: the task, the rules, the input format and the required output structure. In production systems prompts are versioned artefacts, tested against real cases like any other code.