Big improvement from V3!
#2
by
pot99rta
- opened
I'm really liking the coherent responses from this model. The biggest issues it seems to have with complexed prompting involves adding {{char}} in the prompt.
It seems to mix characters in the group chat a lot - (thinking they're the same character) and slight a misuse of certain character descriptions. But other than that this is the MOST solid low billion parameter model I've seen, this is the quality all 12 - 16B models should be achieving which gives me some hope. I'm sure a V5 would be godlike if you pull off some of your coding magic~
This review is based off of the 5 quant M. I'm sure the quality would be even better if I ramped it up so it shows how good this model is!