Definition
Fine-tuning continues training from pretrained weights on data chosen for a target task, domain, behavior, or output format. Full fine-tuning updates most or all model parameters. Partial methods freeze some layers or train a small set of added parameters. Low-rank adaptation is one parameter-efficient method, not a synonym for fine-tuning as a whole.
The new data can be labeled examples, instructions and responses, preference pairs, domain text, or another task-specific signal. The objective matters as much as the data. Supervised instruction tuning, continued language-model training, and DPO all change weights, but they optimize different targets and should be named separately when that distinction affects evaluation.
OpenAI's 2018 generative-pretraining work and Google's 2018 BERT work are influential examples of pretraining followed by task fine-tuning. Both showed that one pretrained Transformer could be adapted across several language tasks with limited architecture changes. They popularized the modern workflow, but fine-tuning as a term and practice predates those models.
Fine-tuning can improve a target metric while narrowing other capabilities, worsening calibration, or overfitting a small dataset. Evaluation should therefore cover the intended task, retained base capabilities, safety behavior, and production inputs that differ from the tuning set.
Distinguish it from nearby terms
Fine-tuning changes model parameters. Prompting and retrieval change runtime context. Tools add external actions or information. Distillation trains a student to imitate a teacher, while fine-tuning adapts the selected model to new data or objectives.
Check your understanding
A fine-tuned support model follows the company's answer format but loses accuracy on uncommon product versions. Which retained-capability tests should have run alongside the target-format evaluation?