Fine-Tuning
Fine-tuning is further training of an already-trained language model on a smaller set of example inputs and outputs, so that it answers in a particular format, tone or domain. It changes the model’s weights; it does not give it access to anything new at answer time.
What changes inside the model
A base model has already learned from a large general corpus. Fine-tuning continues that training on a narrow dataset, in OpenAI’s guide a JSONL file of example prompts and the responses wanted for them. The guide lists what that buys: consistent formatting for a domain, better handling of unusual inputs, shorter prompts at scale, and the ability to make a smaller model perform a specific task well. It also recommends trying prompting first, and treats fine-tuning as the step after that.
What fine-tuning does not do is keep the model current. Its knowledge still ends where the training data ends. Facts that change, such as a price, an opening time or whether a business exists, reach an answer through retrieval: the assistant searches, fetches pages and reads them while it writes. That route runs on what a website serves today, and it is the route generative engine optimization acts on.
A common question
Can a business fine-tune ChatGPT so it recommends them?
No. Fine-tuning is something a developer does to a copy of a model through an API, using their own examples, and the result is a model that developer runs. The public assistants a customer talks to are not fine-tuned on request, and nothing a business puts on its website enters a model this way. What a business controls is what its pages serve when the assistant fetches them, which is the retrieval route, not the training route.
Related terms
Sources
Mechanism and recommended use: OpenAI, fine-tuning guide. No figures on this page are this site’s own.