Technique
Adapter
Small trainable module inserted into a frozen base model for parameter-efficient tuning.
Definition
Adapters add tiny bottleneck modules — typically a down-projection, non-linearity and up-projection — inside each transformer layer. Only adapter weights are trained, leaving the base intact. LoRA is the dominant variant today, but earlier adapter designs (Houlsby, Pfeiffer) work similarly.
Common use cases
- Multi-task serving
- Per-tenant tuning
- Domain adapters