mlabonne/Marcoro14-7B-slerp

Marcoro14-7B-slerp: A High-Performing 7B LLM

Marcoro14-7B-slerp is a 7 billion parameter language model created by mlabonne through a merge of two base models: AIDC-ai-business/Marcoroni-7B-v3 and EmbeddedLLM/Mistral-7B-Merge-14-v0.1. This merge was performed using the slerp method, a technique for interpolating between model weights.

Key Capabilities & Performance

This model stands out for its strong benchmark performance, positioning it as a top-tier 7B model on the Open LLM Leaderboard. It demonstrates significant improvements over models like OpenHermes-2.5-Mistral-7B across several evaluation suites:

• Open LLM Leaderboard: Achieves an average score of 73.01, with notable results in AI2 Reasoning Challenge (69.80), HellaSwag (87.13), MMLU (65.11), and GSM8k (70.89).
• Nous' Benchmark Suite: Scores an average of 57.67, outperforming OpenHermes-2.5-Mistral-7B by over 5 points. Specific gains include +1.59 on AGIEval, +3.12 on GPT4ALL, +11.11 on TruthfulQA, and +4.68 on Bigbench.
• Reasoning: Shows strong capabilities in reasoning tasks, as indicated by its AGIEval and Bigbench scores, including logical deduction and causal judgment.
• Truthfulness: Achieves a TruthfulQA mc2 score of 64.15, suggesting a good ability to generate factually correct responses.

Good For

• Applications requiring a highly capable 7B model for general-purpose text generation and understanding.
• Tasks that benefit from strong reasoning and analytical skills.
• Developers looking for a performant and efficient model in the 7B parameter class.