Model Overview

L3.3-Damascus-R1 is a 70 billion parameter language model developed by Steelskull, utilizing a custom DeepSeek R1 Distill base named Hydroblated-R1-V3. It employs the novel SCE (Select, Calculate, and Erase) merge method to ensure stability, intelligence, and enhanced awareness. The model integrates specialized components from several high-performance models, including:

• EVA-LLaMA-3.33 for core capabilities and creative expression.
• Euryale-v2.3 and Cirrus-x1 for enhanced reasoning.
• Anubis-v1 for detailed scene description.
• Negative_LLAMA for reduced bias and balanced responses.

Key Capabilities

• Enhanced Reasoning: Integrates elements from Euryale and Cirrus for improved logical coherence.
• Creative Expression: Leverages EVA foundations for strong creative writing and scene comprehension.
• Detailed Generation: Incorporates Anubis components for rich and descriptive outputs.
• Balanced Perspective: Includes Negative_LLAMA to mitigate bias and promote balanced responses.
• Stability and Coherence: Built with a custom base and SCE merge method, focusing on reliable performance.

Good for

• General Use: Designed for a broad range of applications.
• Roleplay and Story Generation: Specifically optimized for creative writing tasks, including roleplay and detailed narrative creation.
• Natural Conversation Flow: Aims to provide intelligent and engaging interactions while maintaining a natural conversational style.

Benchmarks

As of February 2, 2025, L3.3-Damascus-R1 achieves a UGI Score of 36.5, a Willingness Score of 2.5/10, and a Natural Intelligence score of 45.62. Its Political Lean is -17.8% with a Liberalism ideology. Aggregated scores show strong performance in Diplomacy (65.5%) and Society (60.0%).