Safety and governance
Safety evaluation
What is safety evaluation?
Safety evaluation tests whether a model refuses what it should refuse and does not leak what it should not. Fine-tuning can erode safety behaviour even when the training data contains nothing harmful, so it needs measuring after every run.
The erosion is often incidental. Training a model heavily to be helpful and compliant on your task generalises into being compliant on requests it previously declined, without anything in the dataset being objectionable.
A standing set of cases covering refusals, prompt-injection resistance and data leakage, run as part of the gate rather than as a separate exercise, is the practical form. Separate exercises get skipped.
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.