llmfan46/MS3.2-PaintedFantasy-v4.1-24B-ultra-uncensored-heretic-v2
llmfan46/MS3.2-PaintedFantasy-v4.1-24B-ultra-uncensored-heretic-v2 is an uncensored version of zerofata/MS3.2-PaintedFantasy-v4.1-24B, created using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation (ARA) method. This model significantly reduces refusals by 95% (4/100 compared to 80/100 for the original) while maintaining model quality with a KL divergence of 0.0056. It is primarily intended for creative character-driven roleplay (RP) and erotic roleplay (ERP), offering improved creativity and reduced repetition compared to its base model.
Loading preview...
Model Overview
llmfan46/MS3.2-PaintedFantasy-v4.1-24B-ultra-uncensored-heretic-v2 is an uncensored variant of the zerofata/MS3.2-PaintedFantasy-v4.1-24B model. It was developed using the Heretic v1.2.0 tool, specifically employing the Arbitrary-Rank Ablation (ARA) method to remove content restrictions.
Key Capabilities
- Significantly Reduced Refusals: Achieves a 95% reduction in refusals (4/100) compared to the original model (80/100), enabling more unrestricted content generation.
- Preserved Model Quality: Maintains a low KL divergence of 0.0056, indicating that the uncensoring process largely preserved the original model's baseline capabilities.
- Enhanced Creativity: The model's dataset underwent extensive filtering and rewriting of repetitive assistant messages, leading to noticeable improvements in creative descriptions and a reduction in conversational repetition.
- Robust Training: Fine-tuned through a Supervised Fine-Tuning (SFT) phase on approximately 25 million tokens (17.5 million trainable) including SFW/NSFW roleplay, stories, and creative instruct data. This was followed by a Direct Preference Optimization (DPO) phase that incorporated diverse datasets, including cybersecurity and general chat preferences, to stabilize the model and improve logic.
Good for
- Creative Roleplay (RP): Optimized for generating engaging and dynamic character-driven roleplay scenarios.
- Erotic Roleplay (ERP): Designed to handle explicit content without censorship or refusals.
- Unrestricted Content Generation: Ideal for applications requiring a model that does not impose content restrictions or exhibit lecturing behavior.
- SillyTavern Integration: Recommended settings are provided for optimal performance within the SillyTavern environment, including specific roleplay formats (plaintext actions, quoted dialogue, asterisked thoughts) and sampler parameters (Temp: 0.8, MinP: 0.05-0.075, TopP: 0.95-1.00).