Fine-Tuning

Adapting a pretrained model to a narrower task or domain with additional training.

Teach a cataloguing convention with examples

A museum wants its assistant to convert rough object notes into a consistent short label. Its training set contains checked input-and-output pairs.

  1. One training pair

    Input: “Etching; maker blank.” Target: “Etching. Maker unrecorded.” The target demonstrates the chosen convention.

  2. One held-out case

    Input: “Lithograph; maker blank.” Test whether the model uses the convention and preserves the new medium without having seen this item in training.

  3. One changing fact

    The gallery displaying the object can change tomorrow. Supply the current location from the collection system when answering a visitor.

What are you trying to teach?

Fine-tuning updates model parameters using additional data. A recurring format or task-specific behavior can be evaluated on records excluded from training. For frequently changing facts, establish a fresh source and test whether the model uses it correctly.

Why Fine-Tuning Is Used

Teams fine-tune models when prompting alone is not enough. Common goals include better domain vocabulary, more consistent formatting, improved task accuracy, stronger style control, or behavior that reflects a company's preferred way of working. Fine-tuning can be especially helpful when the same pattern needs to be repeated many times at scale.

There are multiple approaches. Some projects update many model weights directly, while others use lighter methods that add a smaller number of trainable parameters. The broader idea is the same: guide the model toward a narrower and more useful behavior without retraining the foundation model from scratch.

Follow the idea

Instruction tuning focuses on following a range of requested tasks.

RAG · Retrieval can supply a collection record as it stands today.

Source · Google’s introduction to model tuning

Return to Fine-Tuning in the A–Z glossary