What does LoRA freeze and what does it actually train, and why does that make many task versions cheap?
LLM & AILevel: IntermediateFine-tuning & Alignment
What does LoRA freeze and what does it actually train, and why does that make many task versions cheap?
Hints
- Which parameters change and which stay frozen?
- Separate memory for weights, gradients, optimizer states and activations.
Answer
LoRA freezes the entire base model and trains only a small pair of add-on matrices, B·A, on a few chosen layers. That is typically well under 1% of the weights. Because the base never changes, a finished task is just those tiny adapters, a few MB each. So you can keep dozens of task-specific versions sitting on one shared frozen base instead of storing a full copy of the model per task.
Source and licence: Skillumen — Awesome AI Engineer Interview Questions · CC0-1.0