AI 101

Adapter

Also known as: adapters

Training

An adapter is a small trainable module added to a (usually frozen) larger model to give it a new skill, without retraining the whole thing.

LoRA is the best-known adapter. The pattern is:

  • Frozen base: the big model, untouched.
  • Adapter: a tiny, trainable add-on for one task.
  • Result: one base, many cheap task-specialists.

The design pattern: adapters let you specialize a model *without destroying* the general one. Keep the base, swap the adapters.