Comment by highfrequency

Interesting that for these small models, it is optimal for the embedding parameters to be a huge fraction of the total (170e6/250e6) = 68%!