It looks interesting.
I've only just started testing, but it already looks promising. It is as you write, the model actually has an interesting, slightly dark humor, and does not shy away from vulgarisms (but without exaggeration, they are balanced at the right moment). I noticed that your model is very sensitive to changes in parameters. I'm using Koboltcpp, and I've played around with the temperature settings and Top p, Top-K, Top-A, etc.
Now I have the settings: temperature 0.75, Top-p: 0.92, Top-k:100, Top-A:0.3, Min-p:0.03. I wanted the model to be creative when creating stories according to the instructions. And I arrived at these settings experimentally... but I think it could be chosen better.
Do you have any recommended settings for your model that are good for creating stories?
edit: I also noticed that sometimes the model in the console shows that it is generating text (although nothing is displayed - it stops when it reaches the limit - 512). This does not seem to affect anything, apart from wasting a bit of time - maybe it is the fault of my experimenting with the parameters?
Try top_K - > 40 to 80 ;
Then try temp settings starting and .1 and go up.
I have tested it up to 5 (whole number) .
You may also want to disable "dynamic temp" then change / test top_k changes.
Also - the quant(s) you use will directly impact creative output.
IQ4XS, Q4KS, and Q4KM will all output different results.
Likewise the imat version of Grand Story will also generate different output (deeper, more connected in most cases) than the non-imat version.
Same paras suggestions apply to imat to; as well as "quant" differences.