Temperature

How to tune model temperature with evals instead of assuming zero temperature is always best for consistency.

Temperature

We briefly mentioned this above, but setting model temperature to zero is not an answer to both consistency AND task reliability. Setting a lower temperature than default may be strategic, but you'll likely want to run evals against an annotated dataset to find what the optimal setting is for the model and task combination.