How it works
Instead of updating all the billions of weights in a model, PEFT (Parameter-Efficient Fine-Tuning) freezes most of the network and trains only a small set of additional parameters, such as LoRA adapters or prefixes. This drastically reduces memory and compute cost while retaining much of the quality of full fine-tuning. The core idea is that large models already have enough general knowledge and only need targeted adjustments for a specific task.
Practical example
Fully training a 7 billion parameter model requires expensive GPUs and days of processing. With LoRA, a PEFT technique, you train only a few million parameters in hours on a notebook with a modest GPU, and get a model specialized for code or support. The final result is a small adapter file that combines with the base model at inference time.
What it's for
PEFT democratizes fine-tuning: companies and developers can adapt large models to their domain without data center infrastructure. It is the foundation of much of the specialized models available today in catalogs like Hugging Face, and it allows you to experiment quickly with different tasks without duplicating entire models.
Difference between PEFT and full fine-tuning
In full fine-tuning, all weights are updated, which requires more data, memory and time. PEFT sacrifices a little maximum performance in exchange for efficiency and flexibility, making it ideal for most practical use cases. Adapters complement this approach by adding trainable layers to the frozen model.