Fyi logprobs !== confidence.

If you run "bananas,fishbowl,phonebook," and get {"sponge": 0.76}

It doesn't mean that "placemat" was the 76% correct answer. Just that the word "sponge" was the next most likely word for the model to generate.