Heralax/Augmental-13b
Heralax/Augmental-13b is a 13 billion parameter MythoMax-based model, fine-tuned on a unique, high-quality augmented roleplay dataset derived from human-written text and enhanced by GPT-4. This model excels at generating longer, more descriptive responses and features diverse character personalities, making it ideal for advanced roleplaying applications. It leverages a refined data generation approach that scales beyond single characters and improves writing diversity, supporting a 4096-token context length.
Loading preview...
Augmental-13b: Human-Written, AI-Enhanced Roleplay Model
Augmental-13b is a 13 billion parameter model built upon the MythoMax architecture, distinguished by its innovative "augmented data" training approach. Unlike purely synthetic datasets, Augmental-13b is fine-tuned on a high-quality dataset of over 7,850 roleplay examples, which are human-written (from sources like visual novel scripts) and then significantly enhanced by GPT-4. This method ensures a rich diversity of writing styles and character personalities, moving beyond single-character biases.
Key Capabilities
- Enhanced Response Length and Detail: A unique second GPT-4 pass on the dataset specifically expanded selected lines into much longer and more descriptive responses, making the model excel at generating extended, engaging roleplay outputs.
- Diverse Character Personalities: Trained on multiple distinct characters with a wide range of personalities (e.g., Tsunderes, catgirls), ensuring versatile and nuanced interactions.
- Scalable Data Generation: The underlying data generation process is refined, cheaper, and more scalable than previous methods, allowing for broader application and community contributions.
- SillyTavern Prompt Format: Optimized for the SillyTavern prompt format, facilitating easy integration into existing roleplay setups.
Good For
- Advanced Roleplaying: Ideal for users seeking highly descriptive, long-form, and character-rich roleplay experiences.
- Character-Driven Narratives: Excels in scenarios requiring consistent and varied character portrayals across different personalities.
- Developers Interested in Data Augmentation: The model's unique data generation methodology and available training code (e-p-armstrong/amadeus) offer insights for creating high-quality, shareable, and scalable datasets from existing human-written content.