Pretty good.
Nice model, would be excellent if not for the bad and unnecessary cheesy narration at the end of the messages usually. Stuff like "And at the end, she knew her fate would be changed forever." Very often it seems you get this filler type of narration which ruins immersion in my opinion.
As for the good parts, it took my rather simple character card and played it pretty creatively, filling in the blanks and giving them an engaging personality.
The prose is pretty good, more on the straightforward side. I managed to push this model to about 4k tokens multiple times before heavy repetition and the filler narration kicked in, this is all using Min P sampling with dynamic temperature in Koboldcpp.
Don't know what the prompt format is supposed to be but none worked well except the Noromaid format.
I'm guessing this is Q4_K_M?, would be nice to have Q5_K_M and to also have the model files as I'd like to merge this with a LoRA.
Thanks for the feedback, unfortunately it is a Q4_K_S, I do not have access to the original FP16 files nor any extra quants, due to this model being made on the Kobold mergebox. I'll try to fix the prose issues.
@Herman555 in my Various-Quants folder, I have a V3, I'm hoping that one would be better if you do test it.
Just adding this to echo Herman, ive had fun with this model.
Ive been out of playing with LLMs for a month or two and coming back to find that 7B models have gotten way more interesting is a very nice surprise.
Im going to give the V3 a spin right away.
Thank you for the time im sure you have put into this :D
EDIT: Looking in the various folder i see quite a few models, i realize they are experiments from the description and from the naming convention i guess SS, SmartSunset, and it being the most recent update might be the SultrySilicon V3?
Thanks for the feedback, unfortunately it is a Q4_K_S, I do not have access to the original FP16 files nor any extra quants, due to this model being made on the Kobold mergebox. I'll try to fix the prose issues.
All good, just in my experience I have found that sometimes going from Q4 to Q5 can net a substantial improvement.
@Herman555 in my Various-Quants folder, I have a V3, I'm hoping that one would be better if you do test it.
Will see about testing it, thanks for letting me know. EDIT: Just had a look there but not sure which one is supposed to be the V3.
Thank you for the update, will check it out :)
@Herman555
I have made a separate release for that called Lemonade RP, sorry about that.
I tried a dozen of other 7b and 13b models(include lemonade RP), none of them did as good a job.
@TEtroa So you like this model? compared to my other ones?
@TEtroa I have put the V3 in the Various-Quants folder now, perhaps you'd like that one.
Agreed with Herman good parts opinion, I think the same. This model is very impressive and smart for a Q4_K_S!