Overview

yam-sam-7B is a 7 billion parameter language model developed by mayacinka. It is a product of merging three distinct models: cognitivecomputations/samantha-mistral-7b, CorticalStack/shadow-clown-7B-dare, and yam-peleg/Experiment26-7B. The merge was performed using the dare_ties method, with yam-peleg/Experiment27-7B serving as the base model.

Key Capabilities & Performance

This model demonstrates solid performance across a range of benchmarks, as evaluated on the Open LLM Leaderboard. It achieves an average score of 74.58.

Key benchmark results include:

• AI2 Reasoning Challenge (25-Shot): 70.90
• HellaSwag (10-Shot): 87.92
• MMLU (5-Shot): 65.39
• TruthfulQA (0-shot): 71.30
• Winogrande (5-shot): 83.03
• GSM8k (5-shot): 68.92

Usage

The model supports standard text generation tasks and can be easily integrated using the Hugging Face transformers library. It is configured for bfloat16 dtype and includes int8_mask parameters in its merge configuration.

Good For

• General-purpose text generation and understanding.
• Applications requiring a balance of reasoning, common sense, and language comprehension.
• Developers looking for a merged model with competitive benchmark scores in the 7B parameter class.