Quants generated by a model specific imatrix dataset.

If you want a plug and play roleplay experience this is not the model for you. You will have to iterate your settings.

Extremely sensitive to system prompts and character cards.

Give this model a clear signal as to what you want from it, and it will sing.

Confuse it with conflicting instructions or poorly structured cards and it will reflect that garbage right back to you.

An rp finetune of a merge of 6 genre specific finetunes

Baseline setup:
-NO SYSTEM PROMPT
-Well structured card. Also handles dm/multi-char cards well.
-temp 1, min-p 0.02, top-nsigma 2, rep-pen 1.03, standard DRY.
-use top-nsigma as the primary sampling dial.

BEST PRACTICES

Regenerate the first response a few times until it is close to the style you want.
Edit the first few responses into the structure you personally like. i.e. thoughts -> dialogue -> action, or thought bubbles / status blocks. It will remember after one or two instances.
Talk to it like you want it to talk to you. It's desire to tone/quantity match is very strong.
If the model takes a turn you dont like regenerate.
If regenearation doesn't help, give it a direction "ooc:"
If the model wont respond to ooc: on the first try, prefill it's chat with ooc: and let it complete from there.
If the model is rushing things, either tell it to slow down via ooc: or add in "deliver a slow burn narrative" to your author's note, as system, at depth 6 to 10.

If you feel like you have to have a system prompt use the following:

"You are a masterful roleplayer and an experienced narrator, with the ability to embody multiple characters if required. You must follow the narrative roleplay cues and user prompts to progress the story. Introduce new characters if necessary to pursue that goal."

Examples from Q6_k: