What is parameter-efficient fine-tuning (PEFT)?
Parameter-efficient fine-tuning (PEFT) is the general name for methods that customize a large pre-trained model by training a small fraction of its parameters, or a small set of new ones, while the rest stay frozen. LoRA is the most common PEFT method. Others include adapter layers inserted between existing layers, and prefix tuning and prompt tuning, which learn a set of virtual tokens prepended to every input. All of them produce a small artifact that sits on top of an unchanged base model.
Why parameter-efficient fine-tuning matters
PEFT turns fine-tuning from a research project into a routine engineering task. Training touches so few parameters that a single GPU can do it in hours, the result is a file small enough to version in ordinary storage, and the base model keeps its general ability. A managed service provider can hold one base model and a separate adapter for each client’s ticketing vocabulary and tone, loading the right one per request. A medical billing company can adapt a model to its coding conventions without a cluster. Like LoRA, PEFT shapes behavior and format well but is an unreliable way to add new facts, which belong in retrieval.
What to watch for
- Compare a PEFT adapter against careful prompting first, because many tasks do not need training at all.
- Keep the training examples under the same review as production outputs, since errors in them become behavior.
- Record the base model version each adapter was trained against, because adapters do not transfer across versions.