Parameter-Efficient Fine-Tuning (PEFT) is the general category of techniques — including LoRA and QLoRA — that adapt a model by training only a small subset or small addition of parameters, instead of updating the entire model.
PEFT as a Category
| Technique | Core Idea |
|---|---|
| LoRA | Add small trainable low-rank adapter matrices alongside frozen original weights |
| QLoRA | LoRA combined with a quantized frozen base model, for even lower memory use |
| Prompt/prefix tuning | Train a small set of "virtual tokens" prepended to input, rather than modifying model weights at all |
| Adapter layers | Insert small trainable layers between existing frozen layers |
LoRA and QLoRA are currently among the most widely used PEFT techniques, but the category is broader and continues to evolve — treat specific technique popularity as something that can shift over time.
Why PEFT Exists — The Shared Motivation
Full fine-tuning problems PEFT addresses:
- High compute/memory cost (updating billions of parameters)
- Large storage cost per fine-tuned variant (a full model
copy per task/customer)
- Risk of "catastrophic forgetting" — degrading the base
model's general capabilities while over-specializing on
the fine-tuning data
PEFT's shared approach:
- Keep most/all of the original model frozen
- Train only a small number of additional parameters
- Get most of the benefit of fine-tuning at a fraction of
the costPractical Use Case
PEFT techniques are now the default starting point for most practical fine-tuning projects outside large, well-resourced labs — full fine-tuning of large models remains relevant in specific scenarios but is a much higher-cost undertaking than most teams need for typical behavior-adaptation goals.
Common Mistakes
- Assuming "fine-tuning" always means full fine-tuning — in practice, most real-world fine-tuning today uses a PEFT technique
- Choosing a specific PEFT technique based on outdated information rather than checking current, actively maintained tooling and guidance
Interview Relevance
"What's the general category of technique that LoRA belongs to, and why does that category exist?" — Parameter-Efficient Fine-Tuning (PEFT), existing to reduce the compute, memory, and storage cost of adapting large models compared to full fine-tuning.
Practice Question
Explain, in your own words, the shared underlying goal connecting LoRA, QLoRA, and other PEFT techniques.
