NeuralDarewin-7B Overview

NeuralDarewin-7B is a 7 billion parameter language model developed by mlabonne, created through a merge of several high-performing Mistral-based models using the dare_ties merge method. This approach combines the strengths of models like Intel/neural-chat-7b-v3-3, openaccess-ai-collective/DPOpenHermes-7B-v2, fblgit/una-cybertron-7b-v2-bf16, openchat/openchat-3.5-0106, OpenPipe/mistral-ft-optimized-1227, and mlabonne/NeuralHermes-2.5-Mistral-7B.

Key Capabilities & Performance

This model demonstrates robust performance across a range of benchmarks, as evaluated on the Open LLM Leaderboard. It achieves an average score of 71.79, with notable results including:

• AI2 Reasoning Challenge (25-Shot): 70.14
• HellaSwag (10-Shot): 86.40
• MMLU (5-Shot): 64.85
• TruthfulQA (0-shot): 62.92
• Winogrande (5-shot): 79.72
• GSM8k (5-shot): 66.72

These scores indicate strong capabilities in reasoning, common sense, language understanding, and mathematical problem-solving.

Good For

• General-purpose instruction following: Its diverse training base makes it suitable for a wide array of conversational and task-oriented applications.
• Reasoning tasks: Performance on ARC and GSM8k suggests proficiency in logical and mathematical reasoning.
• Language understanding: Strong HellaSwag and Winogrande scores indicate good contextual comprehension.