AWS Certified AI Practitioner AIF-C01 Practice Question

Parameter-efficient fine-tuning (PEFT) techniques-such as adapters or LoRA-freeze the original model weights and learn a small set of additional parameters. Because only a few million parameters are updated and saved, GPU memory, training time, and storage costs are far lower than with full model fine-tuning or pre-training from scratch. In-context learning changes no weights at all, so it cannot permanently adapt the model, and full fine-tuning or training a new model both require updating and storing all 10 billion parameters, exceeding the team's resource limits.