OFA-Sys/gsm8k-ScRel

Is majority voting(self-consistency) helpful for 70B llama2-sft model?

platoonpluto opened this issue · 1 comments

platoonpluto commented
Is majority voting(self-consistency) helpful for 70B llama2-sft model?
GanjinZero commented

We have not enough resources to complete that experiment. From my experience, using a slightly larger temperature (maybe larger than 0.7) will be helpful with self-consistency for a large SFT model.