I wonder if there is any possibility that you can make your collected pairwise dataset open-source ?
#4
by
fakerbaby
- opened
The 0.4B trained pairRM can present great power for ranking and comparing; however, practically, the model's ability can not generlized to some model's outputs. Based on recent papers, I believe big RMs trained on the same data can demonstrate more generalized ability.
fakerbaby
changed discussion title from
I wander if there is any possibility that you can make your collected pairwise dataset open-source ?
to I wonder if there is any possibility that you can make your collected pairwise dataset open-source ?
Please check out processed Unified-Feedback