Back to glossary

Term

Fine-Tuning

Continued training of a pretrained model on a smaller, specialized dataset to adapt style, format or domain knowledge in a targeted way.

Fine-Tuning — in more detail

Fine-tuning is the process of continuing training of an already pretrained model (foundation model) on a much smaller, curated dataset. The goal is not to teach the model brand-new general knowledge, but to sharpen behavior, tone, format or specific tasks. Common approaches range from full fine-tuning (all weights are updated) to parameter-efficient methods like LoRA and QLoRA, which train only a fraction of the parameters.

Example / practical context

A typical use case is a chat assistant expected to respond in a company-specific style with fixed answer structures. Instead of cramming those rules into every prompt, you collect 500 to 5,000 high-quality example dialogues and fine-tune the model on them. The result is shorter prompts, more consistent responses and often cheaper inference, because a smaller fine-tuned model can match a larger untuned one.

Delineation from similar terms

Fine-tuning changes the model weights, RAG only augments the prompt at runtime — frequently changing knowledge belongs in RAG, behavior and style usually in fine-tuning. Pretraining is the resource-intensive baseline run on massive datasets; fine-tuning builds on top of it and is orders of magnitude cheaper.

Entdecke mehr