Fine-tuning
Also: finetuning, fine-tune
Starting from pretrained weights rather than random ones, training continues on labeled or curated data for the target task, updating all weights or only a few. Parameter-efficient methods such as LoRA freeze the base model and train small add-on matrices.
Example
For a 4096×4096 weight matrix (≈16.8M parameters), a rank-8 LoRA adapter trains only 8×4096 + 4096×8 ≈ 65K parameters, under 1% of the original.