Model Overview

Steelskull/L3.3-Nevoria-R1-70b is a 70 billion parameter model developed by Steelskull, built on the Llama 3.3 foundation. This iteration, Nevoria-R1, integrates the DeepSeek-R1 reasoning architecture to significantly enhance dialogue interaction and scene comprehension. It maintains the core strengths of its predecessors (EVA, EURYALE, Anubis) in storytelling and detailed scene descriptions, while also incorporating elements from Negative_LLAMA and Nemotron for a balanced perspective and reduced positive bias.

Key Capabilities

• Enhanced Storytelling: Excels in generating creative narratives and detailed scene descriptions.
• Improved Reasoning: Leverages DeepSeek-R1 for better logical coherence and prompt adherence.
• Dialogue Interaction: Designed for more nuanced and intelligent conversational exchanges.
• Unique Response Characteristics: Utilizes an unconventional "weight twisting" merging method, leading to distinct model behavior, including a tendency to identify and address minor prompt issues.

Performance Metrics

• UGI Score: 42.66 (with a Willingness Score of 2.5/10 and Natural Intelligence of 40.64).
• Open LLM-Benchmark Average: 43.68%, with notable scores in IFEval (60.24%) and BBH (56.17%).

Good For

• Creative Writing & Roleplay: Ideal for applications requiring rich narratives and descriptive content.
• Interactive Dialogue Systems: Suitable for scenarios demanding enhanced conversational intelligence.
• Users Seeking Unique Model Behavior: Its distinct merging approach offers a different interaction style compared to conventionally merged models.