Settings / known issues and fixes.
(this is on the model card page; posted here for clarity)
Settings / Known Issue(s) and Fix(es):
This merge runs very hot, so a reduction in "temp" may be in order for some use cases (IE: ".8" down to ".6").
If you getting a repeat word (IE: "hahaha" , "ahhhh", "f-word"), and/or letter (IE "nnnn") change the "rep penalty" to "1.1" to "1.5"
Set "rep pen" higher if you are getting "repeat" letter(s), lower if you are getting a repeat word(s).
Generally "repeat penalty" setting of "1.1" works well.
Role Play / Multi-turn: Rep_pen of 1.15 or higher is recommended.
I would also like to write about the fact that the model sometimes begins to either confuse or forget where to put the signs *, ", '.
looking into this.
By the way, I would like to know if you are going to make more models? If so, will you make a new model with a larger volume or a slightly smaller one?
Releasing a new version called: Grand Story in 24 hours. This is full f32 precision based on TWO 16.5B llama story models mastered at F32.
This is far above Little LLM and Grand Horror 16.5B in terms of performance, precision and detail - and rock solid stable to boot.