Fine-tuning

Fine-tuning takes a model that already has general language ability and keeps training it on a smaller, focused dataset until it picks up a particular style or task. It adjusts the existing weights instead of starting from random initialization, so it needs a fraction of the data and compute that pretraining did.

Full fine-tuning updates every weight. Adam, the standard optimizer, keeps the weights, the gradients, and two optimizer states resident at once. A 7 billion parameter model that stores as 14 GB of fp16 weights can need on the order of 100 GB to fully fine-tune. Low-rank methods avoid that. LoRA freezes the base weights and trains a small set of added parameters, and QLoRA stacks 4-bit quantization on top. The original QLoRA paper fine-tuned a 65 billion parameter model on a single 48 GB GPU.

Full fine-tuning wants data center memory and usually more than one GPU. LoRA and QLoRA bring the same job within reach of a single high-memory workstation card, which is how nearly all local fine-tuning runs get done.

Sources

Last verified August 29, 2026.