>Refuel provides LLMs that can compute confidence scores for every label, if the LLM you've chosen doesn't provide token-level log probabilities. How does this

>Refuel provides LLMs that can compute confidence scores for every label, if the LLM you've chosen doesn't provide token-level log probabilities. How does this work exactly?