yes_loc is not used?
#10
by
ymotter
- opened
Just trying to understand the logic. The yes_loc is used in FlagLLMReranker, but not in LayerWiseFlagLLMReranker. How is the probability computed?
The FlagLLMReranker retains the original structure of the model such that the final head layer maps to multiple tokens, necessitating the extraction of logits at the position where the 'Yes' token is located. In contrast, the LayerWiseFlagLLMReranker modifies the structure of the original model by keeping only the linear layer in the head that maps to the 'Yes' token, thus the final output consists solely of the logits for 'Yes'.