Marcoroni-7b-DPO-Merge: A High-Performing 7B Model

Marcoroni-7b-DPO-Merge is a 7 billion parameter language model developed by nfaheem, created through a strategic merge of several existing models using mergekit. The base model for this merge is madatnlp/marcoroni-7b-v3-safetensor, combined with fblgit/UNA-TheBeagle-7b-v1 and udkai/Turdus using the TIES merge method.

Key Capabilities & Performance

This model demonstrates strong performance across a suite of benchmarks, achieving an average score of 74.9. Specific benchmark results include:

• ARC: 73.04
• HellaSwag: 88.8
• MMLU: 64.24
• TruthfulQA: 70.47
• Winogrande: 85.24
• GSM8K: 67.63

Notably, as of January 15, 2024, Marcoroni-7b-DPO-Merge was ranked #1 on the HuggingFace Leaderboard among models around the 13B parameter size, outperforming models like mlabonne/Beagle14-7b and udkai/Turdus in overall average score.

Good For

• General text generation: Its balanced performance across various benchmarks suggests suitability for a wide range of generative AI tasks.
• Reasoning and common sense tasks: Strong scores in ARC, MMLU, and Winogrande indicate good capabilities in these areas.
• Developers seeking a highly performant 7B model: Its top ranking on the HuggingFace Leaderboard highlights its competitive edge within its parameter class.