# Prefix tuning

Prepends trainable vectors to the attention keys and values at each layer while the model stays frozen.

## What it is used for

Lightweight task adaptation without changing the model weights.