SubMaroon-exp/Boulesis-v2-26B-A4B
SubMaroon's Boulesis-v2-26B-A4B is a 26 billion parameter Gemma-based model, the second iteration in the Boulesis series. This version is a composite model utilizing QK task arithmetic and fused LoRA, specifically designed to enhance prose diversity, decisiveness, and context retention for roleplay scenarios. It focuses on concisely maintaining logic and deepening character understanding, making it suitable for complex interactive narratives.
Loading preview...
Overview
Boulesis-v2-26B-A4B is a 26 billion parameter model developed by SubMaroon, representing the second iteration in the Boulesis series. It is a composite Gemma-based model that integrates QK task arithmetic and fused LoRA techniques. The primary goal of this version is to refine the model's ability to generate diverse and decisive prose, while significantly improving its attention to context and character details in roleplay (RP) sessions.
Key Enhancements and Differences from v1
- Improved Context and Character Grasp: Version 2 aims to sharpen the model's ability to organically integrate lore and facts from character cards, moving beyond simple mirroring of user input.
- Concise and Logical Output: This iteration is noted for its very concise writing style, maintaining strong logical coherence, and potentially deepening character understanding.
- Training Data: Unlike v1, v2 was trained exclusively on English data.
- Technical Adjustments: Significant changes were made to QK alpha values, LoRA layers (10-28 layers in v2 vs. all 30 in v1), LoRA targets (35 in v2 vs. 55 in v1), and bake scale (0.70 in v2 vs. 0.26 in v1).
Recommended Usage
For optimal performance, users are strongly advised to enable "Thinking" mode. Recommended sampler settings include a temperature of 1.0, Top-K of 64, repetition penalty between 1.05-1.1, and Top-P of 0.95. Specific configurations for KoboldCPP and SillyTavern are provided to ensure reasoning functions correctly.