Adapting a model to a specific purpose.
THE DEFINITION
Taking a trained model and training it further on a smaller, specific dataset.
WHY IT WORKS
Training from nothing requires enormous resources. Adapting an existing model requires far less.
WHAT IT IS USED FOR
Adapting to a specialist vocabulary Teaching a particular output format Aligning to an organisation's tone Improving performance on a narrow task
WHAT IT CANNOT DO RELIABLY
Add knowledge. Fine-tuning shapes behaviour more than it installs facts.
For giving a model access to your information, retrieval works better.
THE ALTERNATIVE MOST PEOPLE NEED
Prompting well. A carefully constructed prompt achieves much of what people expect from fine-tuning, at no cost.
WHEN IT IS WARRANTED
High volume, a consistent specialised format, or a narrow task where a small tuned model beats a large general one.
RELATED TERMS
Training, retrieval augmented generation, prompt.
