Ber666/ToolkenGPT

logits_bias

lisaalaz opened this issue · 1 comments

Hi, I hope you are well! I just have another quick question about the code if I may (I couldn't find this in the paper): how did you tune the various logits_bias hyperparams that you use at inference? Many thanks!!

Ber666 commented

Hi, in our approach, we do not have standardized procedure for tuning the bias (since there is only one single value). Instead it's empirically determined based on observations from a few test cases.