SamChat: A Merged 7B Language Model

SamChat is a 7 billion parameter language model developed by fhai50032, constructed through a merge of two distinct models: Intel/neural-chat-7b-v3-3 and cognitivecomputations/samantha-mistral-7b. The merging process utilized the DARE TIES method, with macadeliccc/WestLake-7B-v2-laser-truthy-dpo serving as the base model.

Key Capabilities & Performance

SamChat demonstrates solid performance across a range of benchmarks, as evaluated on the Open LLM Leaderboard. It achieves an average score of 61.68, with notable results in:

• AI2 Reasoning Challenge (25-Shot): 62.20
• HellaSwag (10-Shot): 81.88
• MMLU (5-Shot): 59.70
• Winogrande (5-Shot): 72.14

These scores indicate its proficiency in general reasoning, common sense understanding, and multi-task language comprehension. The model's configuration includes specific weight and density parameters for its merged components, optimized for float16 dtype.

Use Cases

SamChat is suitable for general text generation and conversational AI applications where a 7B parameter model with balanced reasoning and language understanding capabilities is required. Its merged architecture aims to combine the strengths of its constituent models for improved overall performance.