Is I-* variant models supposed to be better or worse than non I-* variant?
#1
by j5IGNPff2vG - opened
title
the I-* Variants uses I-Matrix. These perform better in eval benchmarks such as Winogrande and HellaSwag, but they are slightly worse in core metrics such as Perplexity and KD (but, neglectable). I'd suggest to use I-*, these are qualitatively much better.