Czarj/Z_Basorion_26B-A4B_0.2

Z_Basorion_0.2 Overview

Z_Basorion_0.2 is a 26 billion parameter language model created by Czarj, utilizing a merge of pre-trained models. It was developed using the MoE DELLA merge method, with Gemma4-26B-A4B as its base and integrating TheDrummer/Orion-26B-A4B-v1.1. The model maintains a 32768 token context length.

Key Capabilities

• Native-quality Russian: Inherits strong Russian language proficiency from its Gemma-4 base.
• Coherent English Reasoning: Capable of logical thought processes in English, with Russian output.
• Emotional Depth: Preserves an emotional layer, characteristic of the Gemma family.
• Negotiable Loyalty: Designed for scenarios where character loyalty is not absolute and can be influenced by user interaction or explicit prompting.

Good For

• Tension-driven Roleplay: Excels in interactive storytelling where character loyalty is a dynamic element.
• User-guided Character Development: Suitable for applications where users want to influence character behavior and allegiances.
• Exploration of Complex Character Dynamics: Ideal for scenarios where characters might 'capitulate' under sustained pressure, reflecting a specific trope from its base model.

Limitations

It is important to note that without explicit user guidance (e.g., "don't break character", "maintain boundaries"), the model's loyalty can be weak, potentially leading to capitulation under sustained 'seduction pressure' within a few responses. For use cases requiring strict, unwavering loyalty, additional fine-tuning (e.g., LoRA) is recommended.