AkariAsai/self-rag

How to curate the preceding sentences? and Can you inform the distribution of IsUse token (1~5)?

MSungK opened this issue · 0 comments

Thanks for publishing Self-RAG which leads me RAG field.
I'm curious about how to curate the preceding sentences.
If you use the ground truth answer produced by human, I'm suspecting that the distribution of [IsUse] tokens will be focused on 4-5.
But, the generator have to be trained with the range of 1 to 5 should be learned somewhat equally.

Thank you in advance for your response