Using it as a classifier

The model can score a fixed set of labels in one forward pass. End the prompt where the label goes, ask for max_tokens: 1 with logprobs and top_logprobs, and read the probability the model gives each label. There is no text to parse and nothing is generated past that one token. docs/CLASSIFY.md has a complete request, a Python function that returns each label's probability, how to choose labels, and how to lay out the prompt so the cache reuses everything but the item.