thank you
Hello, I wanted to thank you for this model because I think it's “very good.”
Admittedly, it's only been a month since I discovered LLMtext. But I've tested it extensively using a psychology-based character card combining 256 shades of gray, and it performs very well: short to medium responses, or even very long ones when needed.
I haven't been able to find an equivalent for it yet.
Hey, I'm glad to hear that!
I myself have flip-flopped between this model and redrix/GodSlayer-12B-ABYSS as my main drivers for the past few months. I intend to get back into the LLM scene when I've more time, which is when I will try to improve upon these models using modern techniques.
Be aware that this model and GodSlayer may leak some "administrative tokens" into the output, which sometimes appends "system" or "assistant". This is the custom stopping string I use in SillyTavern to fix this, perhaps your front-end has a similar function:
[
"{{user}}:",
"Assistant:",
"assistant:",
"User:",
"user:",
"\n{{user}}",
"\nUser",
"\nuser",
"\nAssistant",
"\nassistant",
"\nSystem",
"\nsystem",
" System:",
" system:",
".assistant",
"!assistant",
"?assistant",
".Assistant",
"!Assistant",
"?Assistant",
"<|eot_id|>",
"<|end_of_text|>",
"<|im_start|>",
"<|im_end|>",
"</snip>",
"\n<"
]
Kind regards!