AI · Flashcard
What is PEFT (parameter-efficient fine-tuning)?
- AFine-tuning only a small subset of parameters instead of all of them
- BFine-tuning every weight, but for far fewer steps than a full run
- CFine-tuning one layer at a time, to keep the optimizer state in memory
- DFine-tuning on data the model generated itself, to avoid labelling cost
Why this is the answer
PEFT adapts a model by training a tiny fraction of its parameters — newly added adapter weights, or a small selected subset — which keeps optimizer state and memory low while matching full fine-tuning on many tasks. Shortening a full run, working layer by layer, or training on self-generated data all still update the whole model, so none of them is parameter-efficient.
Official docs