Gille/StrangeMerges_51-7B-dare_ties

Model Overview

Gille/StrangeMerges_51-7B-dare_ties is a 7 billion parameter language model developed by Gille. This model is a sophisticated merge of five distinct base models, utilizing the dare_ties merging method via LazyMergekit. The constituent models include:

• WizardLM/WizardMath-7B-V1.1: Contributes strong mathematical reasoning abilities.
• Kukedlc/NeuralCoder-7b: Enhances code generation and understanding.
• Weyaxi/Einstein-v4-7B: Provides general language and reasoning capabilities.
• 0-hero/Matter-0.1-Slim-7B-C-DPO: Likely contributes to instruction following and alignment.
• Gille/StrangeMerges_42-7B-dare_ties: An earlier merge from the same creator, suggesting iterative improvement.

Key Capabilities

• Blended Expertise: Combines specialized strengths in mathematics, coding, and general reasoning from its merged components.
• Instruction Following: Benefits from DPO-tuned components, improving response quality and alignment.
• Performance: Achieves an average score of 71.73 on the Open LLM Leaderboard, with notable scores in:
  • HellaSwag (10-Shot): 85.90
  • Winogrande (5-shot): 82.08
  • GSM8k (5-shot): 70.13 (indicating strong mathematical problem-solving)

Good For

• Applications requiring a balance of mathematical problem-solving and code generation.
• General-purpose tasks where robust reasoning and instruction following are important.
• Developers looking for a 7B model with a broad range of capabilities derived from specialized merges.