zerofata/MS3.2-PaintedFantasy-24B

Overview

zerofata/MS3.2-PaintedFantasy-24B is an experimental, uncensored 24 billion parameter model built on the Mistral Small 3.2 architecture, featuring a 32768 token context length. It is specifically designed to excel in character-driven roleplay (RP) and erotic roleplay (ERP), focusing on generating longer, narrative-rich responses with accurate and proactive character portrayals.

Key Capabilities & Training

• Character-Driven RP/ERP: Optimized for detailed and engaging roleplay scenarios.
• Narrative Depth: Produces extended, story-rich outputs.
• Proactive Character Portrayal: Characters are designed to act and react authentically within the narrative.
• Unique Pretraining: Includes a small pretraining phase on light novels and Frieren wiki data, which has shown to improve lore retention.
• Multi-stage Fine-tuning: Underwent Supervised Fine-Tuning (SFT) with approximately 3.6 million tokens, 700 RP conversations, 1000 creative writing/instruct samples, and 100 summaries. This was followed by two stages of Direct Preference Optimization (DPO) to enhance instruction following and mitigate common 'Mistral-isms'.

Recommended Usage

This model is particularly suited for applications requiring highly creative, character-focused, and narrative-intensive text generation, especially within roleplaying contexts. Recommended settings for platforms like SillyTavern include specific formats for actions (plaintext), dialogue ("in quotes"), and thoughts (in asterisks), along with sampler parameters like Temperature (0.8), MinP (0.04-0.05), and TopP (0.95-1.0).