Gille/StrangeMerges_22-7B-slerp
Gille/StrangeMerges_22-7B-slerp is a 7 billion parameter language model created by Gille, formed by merging Gille/StrangeMerges_21-7B-slerp and paulml/OGNO-7B using the slerp method. This model demonstrates strong general reasoning capabilities, achieving an average score of 76.16 on the Open LLM Leaderboard, including 73.72 on AI2 Reasoning Challenge and 64.80 on MMLU. Its primary use case is general-purpose text generation and reasoning tasks, leveraging its merged architecture for balanced performance.
Loading preview...
StrangeMerges_22-7B-slerp Overview
Gille/StrangeMerges_22-7B-slerp is a 7 billion parameter language model developed by Gille, created through a slerp merge of two base models: Gille/StrangeMerges_21-7B-slerp and paulml/OGNO-7B. This merging technique, facilitated by LazyMergekit, combines the strengths of its constituent models to achieve a balanced performance profile.
Key Capabilities
- General Reasoning: Achieves a notable 73.72 on the AI2 Reasoning Challenge (25-Shot) and 64.80 on MMLU (5-Shot), indicating strong logical and multi-task understanding.
- Common Sense & Language Understanding: Scores 89.03 on HellaSwag (10-Shot) and 84.77 on Winogrande (5-shot), demonstrating proficiency in common sense reasoning and pronoun resolution.
- Factuality & Math: Records 74.90 on TruthfulQA (0-shot) and 69.75 on GSM8k (5-shot), suggesting reasonable performance in factual recall and basic mathematical problem-solving.
- Overall Performance: Boasts an average score of 76.16 on the Open LLM Leaderboard, positioning it as a capable general-purpose model in its size class.
Good for
- General Text Generation: Suitable for a wide range of applications requiring coherent and contextually relevant text outputs.
- Reasoning Tasks: Effective for scenarios demanding logical inference and problem-solving, as evidenced by its benchmark scores.
- Balanced Performance: An excellent choice for users seeking a 7B model with well-rounded capabilities across various benchmarks, rather than specializing in a single domain.