Mental BoundFine-tuning takes a model that is already trained and continues its training on your own data, adjusting its internal weights so it specializes in your task, style, or domain while keeping most of what it already knows. It needs far less data and compute than training from scratch — often thousands of examples rather than billions.
Typical uses: a customer support bot tuned on past tickets and tone of voice, a code assistant adapted to a codebase's patterns, a writing tool that learns a brand's style from sample documents.
The main risk is catastrophic forgetting — narrow new data can degrade capabilities the model used to have. Techniques like LoRA, which update only a small subset of weights, help preserve them. For many applications, fine-tuning is the practical route to a model that fits your needs at a fraction of the cost.