Awesome!
Took me a while to figure out the right settings and everything but after that, this model is simply awesome, very smart, and follows directions. Sometimes mixes up characters/words but it's obvious what it is trying to get at so it's a very easy fix, as it seems to seldom switch up words and characters rather than simply be confused with the context completely. Haven't tried it yet at very high contexts yet but so far the writing is significantly better than the previous one too (and in fact it is quite incredibly good). Keep up the great work! I don't think I can foresee replacing or wanting to replace this one with another model for a while.
Good to hear!
Yeah, its working well for me too. It seems to retain good performance up to 45K so far, grasping concepts and details from a big context, and often writing well, but sometimes makes silly mistakes like you said.
I am using 0.05 MinP with 1.25 tau mirostat and 1.08 rep penalty, and adjusting the temperature.
BTW, other models like Sus-Bagel may be quite good if you don't need the mega context.
Nah, a big draw to me is the large potential context. I'm using dynamic temp up to 1.5 with 0.25 min p with 1.18 rep pen. Anyway, I appreciate your efforts!