AI development · AI model fine-tuning
Fine-tune a model only when it beats a good prompt
Fine-tuning trains an existing AI model further on your own examples so it follows a format, style or decision rule without long instructions. It is worth it when you run the same task at high volume and prompting is not consistent enough. We test a strong prompt and retrieval first, and fine-tune only if the numbers show it is better.
What is included
- A comparison of prompting, retrieval and fine-tuning on your task, with results
- Training data collected, cleaned and checked from your real examples
- Training on a hosted model (OpenAI, Gemini) or an open model you can host
- A test set kept apart from training, to show the real improvement
- Notes on cost per request before and after
When is fine-tuning the right choice?
When the task is narrow and repeated often: sorting messages into your categories, extracting the same fields from documents, or writing in one fixed format. When the model needs new facts, retrieval from your documents is usually better and cheaper to keep up to date.
How much data does it need?
It depends on the task. A format or style can be learned from a modest set of good examples; harder decisions need more. We look at what you have before promising anything, and say so if it is not enough.
Questions about ai model fine-tuning
Will a fine-tuned model know our company information?+
Not reliably. Fine-tuning teaches behaviour, not facts. For answers from your documents we use a knowledge assistant with retrieval instead.
Who owns the fine-tuned model?+
It is trained in your account, on your data, and stays yours.