Start with what exists
Pre-trained models already handle a wide range of language and vision tasks. Clear instructions, good examples in the prompt, and the right context retrieved from your own systems solve most business problems without training anything.
When fine-tuning helps
Fine-tuning adjusts a model’s weights using your own labelled examples. It tends to pay off when:
- The task is narrow and repeated at high volume.
- The output format or style must be very consistent.
- The domain looks different from public data, such as a specific camera setup or document type.
- You need a smaller, cheaper model to match a larger one on one job.
| Pre-trained + context | Fine-tuned | |
|---|---|---|
| Time to first result | Days | Weeks |
| Data needed | A few good examples | Hundreds to thousands, labelled |
| Changes | Edit instructions or context | Retrain and re-evaluate |
| Best for | Most workflows | Narrow, high-volume tasks |
What it costs
Fine-tuning needs a representative, labelled dataset, an evaluation set kept separate from training, and a plan to retrain when conditions change. Measure the pre-trained baseline first so you know whether the extra work actually improved anything.