Log-likelihood of output token
#31
by robdei - opened
I can run Pixtral successfully and I receive reasonable responses to my questions.
However, I haven't found documentation on how to determine the log-likelihood of the tokens that are being output by the model. I'm mostly using Pixtral for image classification, so this would be very useful.
Does anyone know where some documentation can be found on this?
Nevermind I think I figured it out.
when callingout_tokens, _ = generate([tokens], model, images=[images], max_tokens=256, temperature=temp, eos_id=tokenizer.instruct_tokenizer.tokenizer.eos_id)
the _ is the logprobs of the tokens.
robdei changed discussion status to closed