Model Overview

aifeifei798/llama3-8B-DarkIdol-2.1-Uncensored-32K is an 8 billion parameter Llama 3-based model, developed by aifeifei798, specifically designed for uncensored and diverse role-playing scenarios. It leverages the winglian/Llama-3-8b-64k-PoSE as its underlying base model, theoretically supporting a 64K context, though tested up to 32K. The model has undergone a new iterative manufacturing process to reduce overfitting and improve stability, making it suitable for various roles and adapted for mobile use.

Key Capabilities

• Uncensored Responses: Provides unrestricted content generation, particularly for creative and imaginative scenarios.
• Role-Playing Specialization: Highly optimized for a wide array of role-playing contexts, including those beyond conventional imagination.
• Quick Response & Detailed Output: Engineered for fast interactions while also capable of generating extensive, thesis-like responses.
• Multimodal Support: Features vision functionality when used with compatible tools like Koboldcpp and the specified mmproj file.
• Mobile Adaptation: Module combination readjusted for better performance on mobile devices.

Good For

• Creative Role-Playing: Ideal for users seeking an uncensored model for imaginative and detailed character interactions.
• Offline Use Cases: Compatible with local inference tools such as Koboldcpp, LM Studio, and Layla, enabling private, offline AI experiences.
• Multilingual Interactions: While primarily tested for Korean and Japanese (with some caveats for Japanese furigana), it aims for broad language support.
• Developers Experimenting with Merged Models: Built using a complex merge of several Llama 3-based models, offering a unique blend of capabilities.