Hugging Face Overview
Model Overview
zerofata/MS3.2-PaintedFantasy-v2-24B is a 24 billion parameter uncensored creative model specifically designed for character-driven roleplay (RP) and erotic roleplay (ERP). This second version significantly improves upon its predecessor by focusing on reducing conversational repetition and enhancing instruction following, aiming for a more dynamic and engaging interaction.
Key Capabilities & Features
- Creative Writing Style: Exhibits a distinct and creative writing style, making interactions more unique.
- Reduced Repetition: Version 2 places a heavy emphasis on minimizing repetitive phrases and dialogue across conversations.
- Improved Instruction Following: Enhanced ability to adhere to user instructions, leading to more consistent roleplay scenarios.
- Uncensored Content: Intended for uncensored creative applications, particularly in RP/ERP contexts.
Training Process
The model underwent a multi-stage training process:
- Supervised Fine-Tuning (SFT): Initial training with RP/ERP, stories, and in-character assistant data.
- Direct Preference Optimization (DPO): Focused on reducing repetition, correcting misgendered characters, and eliminating "slop" (unwanted or low-quality text).
- Kahneman-Tversky Optimization (KTO): Further refined the model to reduce repetition and slop, building on DPO improvements.
Recommended Usage
For optimal performance in roleplay scenarios, specific SillyTavern settings are suggested:
- Roleplay Format: Actions in plaintext, dialogue in quotes, thoughts in asterisks.
- Sampler Settings: Recommended temperature of 0.5-0.6, MinP 0.1, TopP 0.95, and Dry 0.8, 1.75, 4.
- Instruct Template: Utilizes the Mistral v7 Tekken instruct format.
While offering a unique creative flair, the model may occasionally exhibit "brain farts" or inconsistencies due to its specialized nature.