Full fine-tuning

Updating every weight in a pretrained model on a new dataset, producing a complete new set of model weights rather than a small adapter. It is the method with the most capacity to change model behavior, and the most expensive to store, train, and serve: one full copy per task.

Why exams ask this

Tested as the expensive baseline a lighter-weight method gets compared against. The distractor picks full fine-tuning as the default answer for a task that only needed in-context learning or a LoRA adapter.

Relevant to

Related concepts

Resources

No resources linked to this concept yet.