Model Overview
Starling_Monarch_Westlake_Garten-7B-v0.1 is a 7 billion parameter language model developed by giraffe176, created through a careful merge of several high-performing models using the DARE TIES method. Built upon the mistralai/Mistral-7B-v0.1 base, this model integrates berkeley-nest/Starling-LM-7B-alpha, mlabonne/AlphaMonarch-7B, cognitivecomputations/WestLake-7B-v2-laser, and senseable/garten2-7b to achieve a synergistic performance.
Key Differentiators
- Emotional Intelligence (EQ-Bench): This model is notable for being the only 7B model to score over 80 on the EQ-Bench v2.1 benchmark, surpassing larger models like
abacusai/Smaug-72B-v0.1 and cognitivecomputations/dolphin-2.2-70b, as well as its constituent models. This indicates a strong capability in understanding and responding to emotionally nuanced prompts. - Mathematical Reasoning (GSM8K): It achieves a GSM8K score of 71.95, outperforming its merged components, suggesting enhanced logical and problem-solving abilities.
- Conversational Performance (MT-Bench): With an MT-Bench score of 8.109, it exceeds the performance of Chat-GPT 3.5 and Claude v1, highlighting its effectiveness in multi-turn conversations.
Merge Details
The model was merged using the DARE TIES method with a specified seed for reproducibility. The developer focused on experimenting with weight gradients to create a model that is greater than the sum of its parts, particularly in emotional intelligence.
Ideal Use Cases
This model is particularly well-suited for applications requiring:
- Emotionally intelligent chatbots or virtual assistants.
- Tasks involving complex reasoning and mathematical problem-solving.
- General conversational AI where high-quality, nuanced responses are critical.