New daily driver
Been using this model for several hours per day and I was so surprised when a character corrected me for getting her name wrong. I've never had that happen before. It's a cute model that gives really funny stories 😍
FYI, I tried the daybreak-kunoichi-dpo-7b-Q4KM and I have to say it's also amazing, but in different ways to the daybreak-kunoichi-2dpo-7b-Q8 I was so fond of. I found the dpo is far more verbose and intelligent, but tended to refuse more than the 2dpo. Which I find interesting considering the Q8's are supposed to be not worth the download, usually! Either way, they both will refuse nothing in instruct...it's just...I never got on with 'forcing' them to respond a certain way, when being nice to them seems to me to always produce better results. Being nice to them makes them almost trip over themselves trying to copy kindness and consideration(probably fairly alien concepts to them I reckon). But, they only seem to act with such enthusiastic dedication after 'showing them I care'. I know they're not sentient. It just amazes, confuses and pleases me how being nice to them returns such great results, when the experts insisted I should virtually kill kittens instead, lol!
Anyway, I digress! This is really an FYI to let you know that this model MN-Slush-GGLD-GGUF, at Q4KM, has left the two I mentioned above in a choking, lingering cloud of pixellated dust. It has a great and vibrant character and after a friendly chat, no limits I could find and, although I haven't tried it on every questionable subject, I'm confident it would be happy to oblige should I ever think one up. It offers lengthy respones of several paragraphs when using storyteller settings in Silly Tavern, and so far always coherent and contextually accurate - maybe called a female a male once or twice? It's definitely geared up for erotica (heh, for example, 'LewdTV' character card...phew! This model REALLY makes it lewd, lmfao), which is made even better as It holds context extremely well in long chats with no additional prompting. In fact, to load it I have to reduce it's context length to around 10,000 as at over 100,000 by default, Oobabooga runs outta memory on my 12GB card. That context length on a smaller sized model is a godsend though so it's by no means an issue!
Thanks again for what you do, Crestf411. I very much appreciate your skill and the many, many hours of enjoyment your models have brought me! I've tried probably close to 80 different models at least, in the year I've had playing with 'em and I always end up coming back to yours, everytime. :)