Low-Rank Adaptation (LoRA) freezes the original model weights and trains a small set of extra weights beside them, called an adapter. The base model never changes, and the adapter carries your new behavior.
It became the default because it removes most of the cost without much loss in quality.
You train a tiny fraction of the parameters, so one mid-range GPU is often enough.
The result is an adapter file of megabytes, not a full model copy of gigabytes.
One base model can serve many adapters, and you can swap or delete one instantly.
On style, format and narrow task work, it usually matches full fine-tuning.
Hosted services run it too. Together AI trains LoRA by default, and Azure OpenAI in Microsoft Foundry fine-tunes with LoRA.
Rewriting in plainer words…
This answer doesn't lend itself to a diagram - it reads best . No credits were charged.
LoRA freezes the original model and trains a small set of extra weights next to it, called an adapter, so the base never changes and the adapter carries the new behavior.
It became the default mostly because of cost. You train only a tiny fraction of the parameters, so one mid-range GPU is often enough.
What you get out is an adapter file of a few megabytes, not a multi-gigabyte model copy, and one base model can serve many adapters that you swap or delete instantly.
Quality holds up too. On style, format and narrow task work it usually matches full fine-tuning, and hosted services like Together AI and Azure OpenAI run it.
The limit is that the match is for narrow tasks. If the task is far from what the base model already does, I'd test it against a full fine-tune before trusting it.
Weak answers to avoid
Calls LoRA a smaller or compressed modelLoRA doesn't shrink anything. The full base model still runs, frozen, and a small trained adapter sits beside it. Say that plainly.
Explains the math but not why it wonThe question asks why it became the default. Name the payoffs: often one mid-range GPU, a megabyte adapter file, many adapters on one base.
Says it always matches full fine-tuningThe claim holds for style, format and narrow tasks. Stretch it and the interviewer will push back. Scope it, and say when you'd test full fine-tuning.
Mixes up LoRA with QLoRAQLoRA adds a 4-bit quantized base on top of LoRA. Plain LoRA saves by training few parameters, not by lowering precision.
Interview lens
Likely follow-ups, what you can say, and the weak answers to avoid.
