Retreatcost/Shisa-K-sakurization

Retreatcost/Shisa-K-sakurization: An Experimental Roleplaying Merge

Retreatcost/Shisa-K-sakurization is a 12 billion parameter language model created through an experimental merge using mergekit. This model specifically targets enhanced roleplaying capabilities by integrating a LoRa adapter derived from PocketDoc/Dans-SakuraKaze-V1.0.0-12b into a Shisa-K-12B base.

Key Capabilities

• Enhanced Roleplaying: The primary focus of this merge is to significantly improve the model's ability to engage in and generate rich, detailed roleplay scenarios.
• Extended Context: With a 32768 token context length, the model can maintain longer, more complex narratives and character interactions.
• ChatML Format: The model is designed to be used with the ChatML format, facilitating structured conversational inputs.
• Japanese Symbol Output: Users may occasionally encounter Japanese symbols in the output, a characteristic that can potentially be mitigated by adjusting TOP_P to 90 and MIN_P to 0.1.

Merge Details

The model was constructed using the Linear merge method, with ./retokenized_SHK as the base and a LoRa adapter (./lora_Dans-SakuraKaze-V1.0.0-12b-64d) applied with a weight of 1.0. The tokenizer source is Retreatcost/KansenSakura-Radiance-RP-12b.

Good For

• Developers and users seeking a specialized model for creative and immersive roleplaying applications.
• Scenarios requiring deep character engagement and extended narrative coherence.
• Experimentation with merged models for specific domain enhancements.