Fine-tuning is extra training for an AI model that already exists, using a smaller set of examples chosen for one job. The model keeps what it learned before and gets better at a particular task or writing style. For example, a food bank could fine-tune a model on its past newsletters so new drafts sound like the food bank. Fine-tuning costs far less than training a model from scratch, which is why organizations usually start from a foundation model.
Why it matters
A fine-tuned model can absorb details from the examples it was trained on, and those details may appear later in its answers. Suppose you fine-tune on donor records, client case notes or staff emails. You should assume some of that information could come back out. That risk is called training data extraction. Before you give personal or confidential data to a vendor for fine-tuning, ask three questions: who can use the finished model, where it is stored, and how it is protected. Also ask who checks the training examples, since tampered examples are a route for data poisoning.
Published Updated