NexoNimbus-7B: A Merged 7B Language Model
NexoNimbus-7B is a 7 billion parameter language model developed by abideen, created through a strategic merge of two distinct models: abideen/DareVox-7B and udkai/Garrulus. This merging approach leverages the strengths of its constituent models to deliver enhanced performance.
Key Capabilities & Performance
NexoNimbus-7B stands out on the Open LLM Leaderboard, ranking as the 5th best-performing 7B LLM. It achieves an impressive average score of 73.5% across a suite of benchmarks, indicating strong general reasoning and comprehension abilities. Specific benchmark results include:
- ARC Challenge: 68.25% (acc)
- HellaSwag: 70.86% (acc)
- GSM8K: 70.35% (acc)
- Winogrande: 84.84% (acc)
- MMLU: 64.69% (acc)
Additionally, it shows solid performance on TruthfulQA, with an mc2 score of 62.42%, suggesting a good capacity for generating factually accurate responses.
Configuration and Usage
The model was constructed using a slerp merge method, combining layers from both base models. It supports bfloat16 dtype for efficient inference. A Colab notebook is provided for easy deployment and experimentation, demonstrating how to run NexoNimbus-7B in 4-bit precision on a free T4 GPU.
Good for
- General-purpose text generation and understanding tasks.
- Applications requiring strong performance in reasoning, common sense, and mathematical problem-solving.
- Developers looking for a highly-ranked 7B model for various NLP applications.