Model Overview

giraffe176/WestMaid_HermesMonarchv0.1 is a 7 billion parameter language model built upon the Mistral-7B-v0.1 base. It was created using the DARE TIES merge method, combining several specialized models including AlphaMonarch-7B, Noromaid-7B-0.4-DPO, WestLake-7B-v2, and a distilled OpenHermes-2.5-Mistral-7B. The merge process involved deterministic density selection, with a density of 0.58 chosen for each component model to optimize EQ-Bench scores.

Key Capabilities & Performance

This model exhibits exceptional performance in specific benchmarks, particularly in conversational and emotional intelligence:

• MT-Bench: Achieves a score of 8.021875, surpassing ChatGPT-3.5-turbo (7.943750) and Claude-1 (7.900000).
• EQ-Bench v2.1: Scores 77.19 (3 Shot, ooba), outperforming ChatGPT-3.5-turbo (71.74) and Claude-1 (76.83), and even some 70B models.
• Conceptual Design: The merge strategy was designed to leverage Westlake and Distilled Open Hermes for initial understanding and thought processes, while Noromaid and AlphaMonarch guide reasoning and conversation.

Benchmarks

While excelling in MT-Bench and EQ-Bench, its performance on the Open LLM Leaderboard shows an average score of 72.62, with specific scores like MMLU at 64.31 and GSM8K at 69.6. On the Yet Another LLM Leaderboard, it achieves an average of 57.42.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

• High-quality conversational AI: Its strong MT-Bench scores indicate proficiency in multi-turn dialogues.
• Emotionally intelligent interactions: Demonstrated by its leading EQ-Bench v2.1 performance.
• Reasoning and nuanced understanding: The merge strategy emphasizes these aspects through its component models.