Hermes-low-tune-3.1: A Merged 7B Language Model

This model, Hermes-low-tune-3.1, is a 7 billion parameter language model developed by nlpguy. It was created using the task arithmetic merge method with teknium/OpenHermes-2.5-Mistral-7B as its base model.

Key Capabilities & Merge Details

The model integrates capabilities from several distinct models, enhancing its overall performance and versatility. The merge incorporated:

• nlpguy/Hermes-low-tune-2
• charlesdedampierre/TopicNeuralHermes-2.5-Mistral-7B
• openaccess-ai-collective/openhermes-2_5-dpo-no-robots
• flemmingmiguel/Mistrality-7B
• beowolx/MistralHermes-CodePro-7B-v1

Each merged component contributed with a weight of 0.2, indicating a balanced integration of their respective strengths across all 32 layers.

Performance Benchmarks

Evaluated on the Open LLM Leaderboard, Hermes-low-tune-3.1 demonstrates solid performance with an average score of 68.31. Notable scores include:

• AI2 Reasoning Challenge (25-Shot): 65.44
• HellaSwag (10-Shot): 84.60
• MMLU (5-Shot): 64.13
• TruthfulQA (0-shot): 53.59
• Winogrande (5-shot): 78.61
• GSM8k (5-shot): 63.46

These results suggest a well-rounded model suitable for tasks requiring general reasoning, common sense, and mathematical problem-solving. Detailed evaluation results are available on the Hugging Face Open LLM Leaderboard.