Fine-tuning
Fine-tuning will let you adapt a model from the catalog to your own data, so it follows your formats, tone and domain vocabulary without having to be told in every prompt.
How we expect it to work
Section titled “How we expect it to work”- Prepare a dataset of example exchanges and upload it via Datasets.
- Pick a base model from the models marked as fine-tunable.
- Start a job, choosing how long to train and how strongly to adapt.
- Evaluate the result against a held-out split before promoting it.
- Call it by its own model ID, using the same API and keys as any other model.
What we expect to hold true
Section titled “What we expect to hold true”- Adapters, not full retrains — cheaper and faster than training from scratch.
- Your tuned model is private to your workspace.
- Training is billed separately from inference, since it consumes GPU time rather than tokens.
- Your training data belongs to you, and is not used to train anything else.
Good candidates
Section titled “Good candidates”Consistent structured output, a house style or tone, domain jargon, or classification where prompting alone is inconsistent.
Fine-tuning is a poor fit for teaching a model new facts — for that, retrieval over embeddings is usually the better tool.