Adapter Fine-Tuning Glossary

Canonical terms for this workspace. A term lands here once it can be used correctly, not when it is first mentioned, so this grows as lessons are earned.

Usage in this workspace

Three words are used loosely across the field in ways that would make later lessons ambiguous, so they are pinned from the start:

Fine-tuning:
Training a small set of added or selected parameters on a task while the base model's weights stay frozen. When all weights are trained instead, this workspace always says full fine-tuning explicitly.
Avoid: training, retraining, teaching the model

Adapter:
The small set of trainable weights added alongside a frozen base model, together with the configuration describing where they attach.
Avoid: fine-tune (as a noun), LoRA (when the method is not specifically LoRA), checkpoint

Quantisation:
Storing weights at lower numeric precision than they were trained in. In this workspace it always refers to the frozen base unless stated otherwise: adapters stay at higher precision, and quantisation for inference is a separate concern.
Avoid: compression, shrinking, optimisation

Terms

Added as lessons establish them.