dfurman/HermesBagel-34B-v0.1
HermesBagel-34B-v0.1 is a 34 billion parameter language model created by dfurman, formed by merging NousResearch/Nous-Hermes-2-Yi-34B and jondurbin/bagel-dpo-34b-v0.2. This model leverages a slerp merge method to combine the strengths of its base models, offering a robust foundation for various generative AI tasks. It demonstrates strong performance across reasoning and language understanding benchmarks, with an average score of 75.15 on the Open LLM Leaderboard, making it suitable for applications requiring advanced comprehension and generation capabilities.
Loading preview...
HermesBagel-34B-v0.1 Overview
HermesBagel-34B-v0.1 is a 34 billion parameter language model developed by dfurman, created through a strategic merge of two powerful base models: NousResearch/Nous-Hermes-2-Yi-34B and jondurbin/bagel-dpo-34b-v0.2. This merge was executed using the LazyMergekit tool, employing a slerp method with specific parameter configurations for self-attention and MLP layers to optimize performance.
Key Capabilities & Performance
This model is designed to offer strong general-purpose language understanding and generation. Its performance has been evaluated on the Open LLM Leaderboard, where it achieved an average score of 75.15. Specific benchmark results include:
- AI2 Reasoning Challenge (25-Shot): 70.56
- HellaSwag (10-Shot): 85.74
- MMLU (5-Shot): 77.38
- TruthfulQA (0-shot): 67.34
- Winogrande (5-shot): 84.61
- GSM8k (5-shot): 65.28
These scores indicate proficiency in reasoning, common sense, language understanding, and mathematical problem-solving.
Good For
- General-purpose text generation: Creating coherent and contextually relevant text for a wide range of applications.
- Reasoning tasks: Its strong performance on ARC and MMLU suggests suitability for tasks requiring logical inference and knowledge application.
- Instruction following: As a merged model incorporating instruction-tuned components, it is likely to perform well in responding to diverse prompts.
- Research and experimentation: Provides a solid base for further fine-tuning or exploring merged model architectures.