📝 PEFT
Description
< What is it? >
- Parameter-Efficient Fine-Tuning (PEFT) adapts a pretrained model while keeping most of its original weights frozen. Instead, it trains a small set of selected or newly added parameters.
- This reduces training memory, optimizer state, and checkpoint storage. One base model can then support many small, task-specific adapters.
Key points
< Common PEFT methods >
| Method | What is trained |
|---|---|
| BitFit | Only the model's bias terms |
| Prompt tuning | Learnable “virtual token” embeddings added to the input |
| Prefix tuning | Learnable prefix vectors injected into every Transformer layer |
| Adapters | Small trainable modules inserted into model layers |
| LoRA | Low-rank weight updates; the original weight matrices stay frozen |
| QLoRA | LoRA adapters trained over a quantized frozen base model |
| IA³ | Small learned scaling vectors for activations |
- LoRA is a common default for LLM adaptation because it offers a strong balance of quality, memory efficiency, and portable checkpoints. See Fine-Tuning for its core idea.
- PEFT lowers the cost of adaptation, but it does not remove the need for high-quality data, validation, and evaluation.