Overview

royallab/Pygmalion-2-13b-SuperCOT-weighed is an experimental 13 billion parameter language model resulting from a weighted merge of Pygmalion-2-13b and Ausboss's Llama2 SuperCOT loras. This merge was performed using a gradient merge script from zaraki-tools, with a specific 50/50 layer weighting (0.51 ratio) applied to the SuperCOT lora, favoring the first iteration of SuperCOT for better performance.

Key Capabilities

• Roleplaying and Conversational AI: Optimized for generating dynamic and contextually rich responses in interactive narrative and character-driven scenarios.
• Experimental Merge: Represents a unique combination of models, with specific layer weighting to balance the characteristics of Pygmalion-2 and SuperCOT.
• Instruction Format Flexibility: Supports both Metharme and Alpaca instruction formats, allowing for versatile integration into different applications.

Limitations and Considerations

• Bias and Risks: Inherits biases from its base models and niche roleplaying forums, making it unsuitable for factual information or advice.
• Experimental Nature: As an experimental merge, performance may vary, and it is not intended for critical applications requiring high reliability or factual accuracy.

Usage

This model is best suited for creative applications such as text adventure games, character-based chatbots, and interactive storytelling where engaging and imaginative responses are prioritized over factual correctness. Users should be aware of its experimental status and inherent biases.