Skip to main content

Finetuning

Finetuning takes a model that has already been trained on broad, general data and continues its training on a smaller, focused dataset. The result is a model that keeps its general abilities but performs noticeably better on your specific task, domain or style.

The analogy is hiring an experienced professional and giving them a few weeks of training on how your company works. They already know their craft; you are adapting that expertise to your particular needs.

When to finetune​

Finetuning is one of several ways to make a model more useful, and it is not always the right one:

  • Prompting changes the instructions you give the model. Quick and cheap, and often enough.
  • Retrieval (RAG) gives the model the right facts at answer time. Best when the issue is missing or changing knowledge.
  • Finetuning changes the model itself. Best when you need a consistent format, tone or behaviour, or strong performance on a specialised task that prompting and retrieval can't reliably achieve.

A good rule of thumb: reach for retrieval when the model lacks knowledge, and for finetuning when it lacks a skill or behaviour.

What you need​

A finetuning job generally requires:

  • A base model to start from.
  • A training dataset of high‑quality examples in a consistent format.
  • A few training parameters that control how the job runs.

The quality and consistency of the dataset matters far more than its size. A small, clean set of examples usually beats a large, noisy one.

Finetuning in Rational AI​

The platform provides a dedicated Fine‑Tuning section where you pick a base model and a dataset, set the parameters and launch a job. When it finishes, the tailored model becomes available in the AI Model Registry, ready to use like any other model.

Additional resources​

  • Fine‑Tuning settings: start and track finetuning jobs.
  • Datasets: build the training data a job needs.
  • RLHF: a specialised form of finetuning that uses human preferences.