Training
Supervised fine-tuning
Also called SFT.
What is supervised fine-tuning?
Supervised fine-tuning trains a model on pairs of inputs and desired outputs, teaching it to reproduce the target response for a given prompt. It is the default and simplest objective, and the right starting point for almost every task where you can write down what a good answer looks like.
The requirement is that you can produce the ideal output. For formatting, tone, classification, extraction and most task-specific behaviour, you can — usually from work your team has already done and approved.
Where it stops working is when the ideal output is hard to write but easy to check. Mathematical derivations and code that must pass tests are the standard examples: verifying is trivial, authoring thousands of ideal examples is not. That asymmetry is what the other objectives exist for.
Data quality dominates data volume here to an extent that surprises people. A thousand consistent, genuinely good examples will beat ten thousand inconsistent ones, because contradictory targets teach the model that the task is ambiguous.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.