Safety and governance
Refusal
What is refusal?
A refusal is the model declining to answer — because the request is out of scope, because it lacks the information, or because answering would be unsafe. It is a learnable behaviour and it disappears if training data contains no examples of it.
Fine-tuning heavily on helpful answers systematically erodes refusal behaviour, even when nothing in the data is objectionable. The model generalises "always be helpful" past the boundary you intended.
Including refusals in the training set is the direct remedy: cases where the correct output is a decline, with the reason. Without them the model will answer everything.
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.