Overview

New-Dawn-Llama-3-70B-32K-v1.0 is a 70 billion parameter model developed by sophosympatheia, built upon the Llama 3 architecture. It is the result of a complex, multi-level SLERP merge of several Llama 3 70B variants, including bosonai/Higgs-Llama-3-70B, Sao10K/L3-70B-Euryale-v2.1, and abacusai/Smaug-Llama-3-70B-Instruct-32K. A key feature is its extended 32K context window, achieved by incorporating layers from the Smaug-Llama-3-70B-Instruct-32K model.

Key Capabilities

• Extended Context Window: Supports up to 32K tokens, enabling longer and more complex interactions.
• Optimized for Creative Tasks: Specifically designed for roleplaying and storytelling, demonstrating improved nuance and understanding in these areas.
• Uncensored Output: The model is uncensored, allowing for a wider range of creative expression.
• Sampler and Prompting Tips: The developer provides detailed recommendations for sampler settings (e.g., Quadratic Sampling, Min-P) and a Llama 3 instruct format prompt template, including adult content instructions for roleplay.

Good For

• Roleplaying: Excels in generating dynamic and engaging roleplay scenarios.
• Storytelling: Produces coherent and descriptive narratives.
• Long-form Content Generation: The 32K context window is beneficial for maintaining consistency over extended creative works.

Comparison

Compared to sophosympatheia's Midnight Miqu, New Dawn is described as "smarter and understands nuances better," while Midnight Miqu might retain an edge in "raw creative juice." The model's uncensored nature and specific optimizations make it a strong candidate for creative applications where nuanced understanding and extended context are crucial.