This wasn't actually deterministic because despite `do_sample: false`, temperature was still taking effect, causing non-deterministic output. I fixed this and also adjusted repetition penalty to be like KoboldAI's Deterministic preset.
After extensive testing, I've adjusted repetition penalty slightly to be the same as simple-proxy-for-tavern's default preset and ooba's LLaMA-Precise settings preset. This fixed some models talking/acting as User.
Groupchats have their CFG defined a little differently. Chat and
global CFGs are publicly accessible within the window. As a compromise,
individual character CFGs can be injected as part of each character's
prompt. However, these CFG settings need to be adjusted in the character's
individual chats which will carry over to the group.
In addition, make this character logic gated under a checkbox to
always prefer chat unless explicitly specified. Negative cascading
is still open, so individual character negatives can be included
at any time.
Signed-off-by: kingbri <bdashore3@proton.me>
This allows for flexibility with global, character, and chat negative
prompts. Combining prompts is very useful for users who want to maintain
a set of global negatives and then add extra layers on top of that. The
ordering is chat -> character -> global tags due to the specificity of
each.
The guidance scale follows the cascade of chat -> character -> global
due to being one number that is set when CFG is fired. If the guidance
scale is 1, nothing happens.
Signed-off-by: kingbri <bdashore3@proton.me>
