IceAbsintheRP-7b: Role-Playing Optimized Language Model
IceAbsintheRP-7b is a 7 billion parameter model based on the Mistral v0.2 architecture, developed by icefog72. It is specifically designed and optimized for role-playing (RP) applications, emphasizing interactive narrative generation and character consistency.
Key Capabilities & Features
- Role-Playing Focus: The model is fine-tuned for generating engaging and coherent role-play interactions, with specific recommendations for use with SillyTavern (ST) and custom rules/lorebooks.
- Extended Context Window: While the model has a theoretical context limit of 32k tokens, it is recommended to use up to 21k tokens for optimal response quality, with performance degradation noted beyond this point.
- Merge Architecture: This model is a merge of several previous Ice-RP variants (Ice0.148-19.10-RP, Ice0.144-15.10-RP, Ice0.143-15.10-RP, Ice0.147-17.10-RP) using the Model Breadcrumbs method, built upon an Ice0.130-16.06 base.
- Quantization Availability: Various EXL2 and GGUF quantizations are available, including 4.2bpw, 6.5bpw, and 8bpw versions, to support diverse hardware configurations.
- Prompting Guidance: The developer provides detailed guidance on configuring prompt content, including specific instructions for 'planning' mechanisms within ST to improve NPC behavior and reduce 'overthinking' common in pure reasoning approaches.
Ideal Use Cases
- Interactive Storytelling: Excellent for creating dynamic and immersive role-playing experiences.
- SillyTavern Integration: Specifically tailored for use with SillyTavern, offering presets and rules for enhanced performance.
- Character Dialogue Generation: Suited for generating consistent and in-character dialogue for NPCs in text-based adventures.