LoRA (Low-Rank Adaptation)
LoRA fine-tunes a model by adding small trainable adapter matrices while keeping the base model frozen.
Why LoRA works
Full fine-tuning updates all weights. LoRA updates a small low-rank change.
text
base weights frozen + small adapter update
This makes training cheaper and easier to store.
Key settings
| Setting | Meaning |
|---|---|
| rank | adapter capacity |
| alpha | adapter scaling |
| target modules | which layers receive adapters |
| dropout | regularization |
| merge | combine adapter with base for serving |
When to use LoRA
Use it for:
- style adaptation
- structured output behavior
- tool-call patterns
- domain classification
- small-team fine-tuning
Do not use it to store changing facts.
Knowledge check
Q1: What stays frozen in LoRA?
The base model weights.
Q2: What does rank control?
Adapter capacity and trainable parameter count.