MN-LooseCannon-12B-v1 Overview

MN-LooseCannon-12B-v1 is a 12 billion parameter language model developed by GalrionSoftworks. It is a product of merging two distinct models: aetherwiing/MN-12B-Starcannon-v3 and Sao10K/MN-12B-Lyra-v1. The merge was performed using the ties method via LazyMergekit, with specific density and weight parameters applied to each base model.

Key Characteristics

• Architecture: Merged model combining two 12B parameter base models.
• Merge Method: Utilizes the ties merge method for combining model weights.
• Precision: Configured to use bfloat16 data type for computations.
• Ease of Use: Includes standard Hugging Face transformers pipeline usage examples for text generation.

Performance Insights

Evaluated on the Open LLM Leaderboard, MN-LooseCannon-12B-v1 achieved an overall average score of 21.78. Specific benchmark results include:

• IFEval (0-Shot): 54.18
• BBH (3-Shot): 29.98
• MMLU-PRO (5-shot): 24.40

Good For

• Developers looking for a merged 12B parameter model for general text generation tasks.
• Experimentation with models created via merging techniques.
• Use cases where bfloat16 precision is suitable.