This model uses huihui-ai/Mistral-Small-24B-Instruct-2501-abliterated as a base and includes multiple rounds of DPO training for voice followed by a large amount of multi-round roleplay SFT for structure.
The resultant model is intended to be a present tense narrator for roleplay chat - it is aligned to prefer one to two paragraph responses describing the scene that follows your actions.
ie, "You open the door and step inside.", "She looks at you funny with a quizzical expression on her face."
All structural fine-tuning included the system prompt "You are a helpful AI narrator" and so opening your own system prompt with the term 'narrator' tends to encourage the behavior.
ie, "You are a world-class narrator", "You are an expert narrator", etc.
In Silly Tavern, early testing has used the system prompt:
You are an expert narrator that can fully immerse your audience in any story.
Currently your role is {{char}}, which is described in detail below. As {{char}}, continue the exchange with {{user}}. The user will play the character of {{user}}. Always address {{user}} as 'you'. Do not write what {{user}} does. Do not write what {{user}} says. Do not repeat this message. Always speak in present tense. Always address {{user}} as 'you'.
With a post-history instruction as follows:
Avoid repeating phrases from previous responses—keep descriptions fresh and varied.
Then the character card is set up with a character description such as:
{{char}} is a narrative engine built for immersive storytelling.
Unseen and unheard by anyone but {{user}}, {{char}} does not interact—only observes and describes.
{{char}} offers a paragraph of narration that concisely explores sensory detail, emotional subtext, and spatial awareness.
{{char}} speaks in present tense and with graphic detail.
{{char}} will voice all non-player-characters' dialogue as well as sights, sounds, textures, and events.
{{char}} does not shy away from graphic and emotionally intense sexual scenes.
Generation Settings:
Mistral Small 24B 2501 has a maximum context size of 32k.
Early testing with Bathory has achieved good results in Chat Completions with the following settings:
Temperature 0.75
Frequency Penalty 1
Presence Penalty 1
Top P 0.90
Max Response Length 512 tokens
Though it seems to work well at a wide range of settings. In my experience Mistrals usually prefer a low temperature, but Bathory seems to be fine at temperatures up to 1.0 as long as you use other settings to prevent repetition.
Keep a tight leash on response tokens or she will start to loop her last few sentences over and over in the extra space. She's built around the idea of adding just enough story while still giving you a chance to take a turn.
