Model Overview
Open_Maid_Samantha_Hermes_Orca_dare_tiesv0.1 is a 7 billion parameter language model developed by giraffe176. It is a sophisticated merge of several instruction-tuned models, built upon the robust mistralai/Mistral-7B-v0.1 base. The merge was performed using the advanced DARE TIES method, which selectively combines the parameters of multiple models to create a more capable and generalized model.
Key Components & Merge Strategy
This model integrates the capabilities of:
teknium/OpenHermes-2.5-Mistral-7BNeverSleep/Noromaid-7B-0.4-DPOOpen-Orca/Mistral-7B-OpenOrcacognitivecomputations/samantha-1.1-westlake-7b
Each component model contributes distinct strengths, with specific density and weight parameters applied during the DARE TIES merge to optimize performance. The int8_mask parameter was enabled, and the model uses bfloat16 for its data type.
Performance Highlights
Evaluated on the Open LLM Leaderboard, this model demonstrates strong general performance with an average score of 67.80. Notable scores include:
- HellaSwag (10-Shot): 85.48
- Winogrande (5-shot): 80.35
- AI2 Reasoning Challenge (25-Shot): 65.87
- MMLU (5-Shot): 64.50
These results indicate a balanced capability across various reasoning, common sense, and language understanding tasks.
Ideal Use Cases
This model is well-suited for applications requiring:
- General-purpose instruction following: Leveraging the instruction-tuned nature of its merged components.
- Conversational AI: Benefiting from the diverse training of models like OpenHermes and Samantha.
- Reasoning tasks: Supported by its performance on benchmarks like ARC and MMLU.
- Development of custom chatbots or assistants: Providing a strong foundation for further fine-tuning.