Better, but not good
#1
by
Nycoorias
- opened
It still has the issue of spewing out nonsense after a short while, Now it just happens on the 6 line rather then the 2.
I only played around with it for a little bit but so far it feels flat out worse then V1 in every aspect.
Update:
It seems to not like Mirostat at all, using Koboldcpp’s traditional sampling method gives me far better results, but it still tends to brake. I will play around with it more tomorrow and report what I find.
Thanks for the report! let me know if you find something helpful for the community so that I can put in the model card
When I tried it in oobabooga with default settings it was giving pretty good consistency