11/ In summary, we have two main ways to evaluate a model outputs:
1. Get the probabilities that some tokens are continuations of the prompt => compare these probabilities together
2. Get the text generation from the model => compare these text generations to possible choices

→ View original post on X — @thom_wolf,