In plain English
Fine-tuning is like giving a well-educated graduate on-the-job training. You start with a capable general model and train it a little more on examples from your own work so it does that job better.
In practice
Fine-tuning helps with consistent style, specialised formats or narrow classification tasks. It's not the best way to teach a model new facts: for knowledge that changes, RAG is usually cheaper and easier to keep current. Try better prompts first.
Under the hood
Fine-tuning continues gradient training from pre-trained weights on task-specific data. Parameter-efficient methods such as LoRA train small adapter matrices instead of all the weights, cutting cost and memory. Risks include overfitting and losing some general abilities.
Example
"We fine-tuned a small model to sort tickets into our 12 categories."