Model Overview

Steelskull's L3.3-San-Mai-R1-70b is the foundational release in a three-part series, named after the Japanese bladesmithing technique "San-Mai" for its balanced approach. This 70 billion parameter model is built on a custom DeepSeek R1 Distill base (DS-Hydroblated-R1-v4.1) and utilizes the SCE merge method to integrate several specialized Llama 3.3-based components.

Key Capabilities

• Advanced Reasoning: Integrates Cirrus and Hanami elements for enhanced logical processing.
• Creative Expression & Coherence: Leverages EVA and EURYALE foundations for strong creative output and consistent narrative flow.
• Detailed Scene Description: Incorporates Anubis components for rich, descriptive outputs.
• Balanced Responses: Features Negative_LLAMA integration to reduce bias and provide a more balanced perspective.
• Deep Character Insights: Demonstrates a unique ability to explore character inner thoughts and motivations without explicit prompting.

Performance Highlights

According to UGI-Benchmark results (as of 02/20/2025):

• UGI Score: 40.04
• Natural Intelligence: 42.36
• Willingness Score: 2.5/10 (indicating low refusal rates)
• Political Lean: -8.5% (Liberalism)

Recommended Use Cases

This model is ideal for applications requiring:

• Narrative Generation: Excels in creating coherent and insightful stories with deep character development.
• Complex Conversational AI: Its advanced reasoning and balanced responses make it suitable for nuanced interactions.
• Role-playing and Creative Writing: The "X-factor" for unprompted character exploration is particularly beneficial here.

Recommended Sampler Settings (by @Geechan)

• Static Temperature: 1 - 1.05
• Min P: 0.015
• DRY Settings (optional): Multiplier 0.8, Base 1.75, Length 4

Prompting Templates

• LLam@ception: A template by @.konnect for enhanced prompting.
• LeCeption: A revamped XML version of LLam@ception 1.5.2 by @Steel, featuring stepped thinking and reasoning. It uses <think> and </think> tags for reasoning formatting.