giraffe176/WestLake_Noromaid_OpenHermes_neural-chatv0.1

Model Overview

giraffe176/WestLake_Noromaid_OpenHermes_neural-chatv0.1 is a 7 billion parameter language model created by giraffe176 through a merge of several pre-trained models. It utilizes the DARE TIES merge method, with mistralai/Mistral-7B-v0.1 serving as the base architecture. The merge incorporates cognitivecomputations/WestLake-7B-v2-laser, NeverSleep/Noromaid-7B-0.4-DPO, teknium/OpenHermes-2.5-Mistral-7B, and Intel/neural-chat-7b-v3-3.

Key Capabilities & Differentiators

• Optimized for Conversational Roleplay: The primary focus of this model is excelling in conversational roleplay scenarios. It has shown strong performance in maintaining story continuity and pacing, even in situations where other models might loop or fail.
• DPO Training for Uncensoring: The model underwent DPO (Direct Preference Optimization) training using a curated version of unalignment/toxic-dpo-v0.2 data, aimed at slightly uncensoring its responses.
• Merged Architecture: By combining multiple specialized models, it aims to leverage the strengths of each component, particularly for its target use case.

Performance Insights

While the merged model shows a strong focus on roleplay, benchmark testing provides a broader performance context:

• MT-Bench Score: Achieved 7.171875 on MT-Bench.
• EQ-Bench v2.1 Score: Scored 65.56 on EQ-Bench v2.1.
• Open LLM Leaderboard: On the Hugging Face Open LLM Leaderboard, the model achieved an average score of 68.86 across various benchmarks, including AI2 (66.72), HellaSwag (85.37), MMLU (64.67), TruthfulQA (51.50), Winogrande (79.72), and GSM8k (65.20).