L3.3-MS-Nevoria-70B: A Merged Model for Enhanced Narrative Generation
L3.3-MS-Nevoria-70B is a 70 billion parameter language model developed by Steelskull, built upon the Llama 3.3 architecture. This model is a unique merge of several specialized components, intentionally designed to create a balanced and highly capable narrative AI.
Key Composition and Differentiators
This model's unique composition is a key differentiator:
- Base Model: Utilizes a lorablated Nemotron-lorablated base, chosen for its unique "weight twisting" effect during merging, which contributes to distinct response characteristics.
- Storytelling & Detail: Integrates EVA-LLAMA-0.1 for strong storytelling capabilities, EURYALE-v2.3 for detailed scene descriptions, and Anubis-v1 for enhanced prose.
- Bias Reduction: Incorporates Negative_LLAMA to actively reduce the common positive bias found in many LLMs, allowing for more realistic and varied narrative outcomes, including darker themes without soft refusals.
Performance and Use Cases
User reviews highlight its exceptional performance in creative writing and roleplay:
- Creativity & Adherence: Praised for insane creativity, perfect character adherence, and dialogue, with a preference for slow-burn narratives.
- Context Management: Demonstrates strong understanding of context and instructions, capable of managing multiple characters and extensive lorebooks (tested up to 110,000 tokens).
- UGI Benchmark: Achieves a UGI Score of 56.75, with a Willingness Score of 7.5/10, and is noted as the highest-ranked 70B model on the UGI-Benchmark as of January 17, 2025. It also shows a Liberalism ideology with a -8.1% political lean.
- Open LLM-Benchmark: Records an average score of 43.92%, with notable scores in IFEval (69.63%) and MMLU-Pro (50.39%).
Recommended Use
This model is particularly well-suited for:
- Creative Writing: Generating rich, immersive, and detailed stories.
- Roleplay: Maintaining character consistency and handling complex interactive narratives.
- Unbiased Content: Producing content that avoids inherent positive biases, allowing for a broader range of emotional and thematic expression.
For optimal performance, users are recommended to explore specific prompt presets like "LLam@ception" by @.konnect.