LLM decision guide
LoRA vs Full Fine-Tuning
LoRA trains small low-rank adapters while freezing the base model; full fine-tuning updates the full parameter set.
Choose LoRA when
- Memory is limited
- Many task adapters share one base
Choose Full fine-tuning when
- Maximum adaptation is worth the infrastructure
- The entire model can be governed and served
Decision table
| Dimension | LoRA | Full fine-tuning |
|---|---|---|
| Trainable state | Small adapters | All weights |
| Memory | Lower | Higher |
| Serving | Base plus adapter compatibility | Separate complete model |
Example
A realistic choice
A team can test several domain adapters with LoRA before justifying a costly full-model run.
Hands-on
Test the decision
Build a memory and deployment budget for both approaches and define an evaluation gate.