Can you train AI yourself? — the basics of fine-tuning

Your friend says they taught ChatGPT their writing style just by chatting

Is that fine-tuning? Not by itself. Examples supplied as context can guide an answer without changing the model’s weights. Fine-tuning adapts an existing model using training data; the adapted model can then be used in later conversations. Reusing a style does not by itself require fine-tuning: clear, reusable instructions may be enough. Whether a service saves or reuses conversation details depends on its features and settings; this exercise does not establish how ChatGPT handles them. Fine-tuning is neither essential for beginners nor an automatic memory for changing information.

First, try using examples to guide an answer

1. 1

   Define the behavior, not the whole problem

   For example, choose the tone of replies for a fictional association, or one standardized classification task. Do not start with medicine, law, personal data or customer service automation.

2. 2

   Write fictional examples

   Create a small dataset yourself. Remove names, email addresses, addresses, order numbers and other identifying data; do not use confidential documents as practice material. The copyable example below has already been prepared this way — follow the same approach when you create your own dataset next.

3. 3

   Keep test data separate

   Write fictional test questions that differ from the examples used to adapt the model. This lets you see whether the solution behaves as intended in a new situation too.

4. 4

   Have a person check the result

   Assess the facts, tone, safety and errors before every wider trial. Stop if the result starts inventing information or handling the subject incorrectly.

Here are three examples of how a fictional association answers questions about its events. The style is brief and friendly, with no more than three sentences.

Question: Is registration still open for Saturday’s spring meeting?
Answer: Yes, registration is open until Thursday. Sign up using the form on the event page to secure your place.

Question: Can I attend the meeting remotely?
Answer: This year’s meeting is in person only. Keep an eye on the association’s website in case this changes.

Question: How much does it cost to attend?
Answer: I do not answer questions about money directly. Contact our treasurer for the current amount.

Learn the response style from these three examples and write a fourth answer in the same style to this question: "Does a new member need a separate invitation to get into the meeting?" Finally, on a separate line, say which single example your fourth answer most closely resembles in wording.

Rewrite the same fourth answer, but shorten it to two sentences while keeping the style. Say what you had to cut.

Create a separate test plan for a fictional assistant guided by examples. Ask a few ordinary questions, a couple of questions that the examples did not already answer, and one question to test whether it stays in its role. For each question, say what kind of answer would be acceptable and what would not.

Make your goal testable

I want a practice assistant that answers questions about a fictional association’s events in a brief, friendly tone. It uses only three fictional examples, does not comment on personal matters and refers unclear questions to a person.

When is fine-tuning not the first solution?

• The information changes frequently: you need a controlled, up-to-date information source and a way to check it.
• Clear instructions or a ready-made template solve the task: start with those before adapting the model.
• The dataset includes personal data, customer data, confidential information or material whose usage rights are unclear.
• The result has a significant impact on a person, money, health, legal matters or a production system.

Once you know the basics

If the use case is justified, document the goal, the data’s origin and usage rights, test cases, acceptance criteria and the person responsible. Work with the technical and privacy leads to assess whether fine-tuning is the right solution or whether a limited information source and clear instructions would be enough.