liminerity/Blur-7b-v1.22 is a 7 billion parameter language model created by liminerity, formed by merging s3nh/Sonya-Panda-7B-slerp, argilla/distilabeled-Marcoro14-7B-slerp, and Weyaxi/MetaMath-OpenHermes-2.5-neural-chat-v3-3-Slerp using the TIES merging method. This model demonstrates a balanced performance across various benchmarks, achieving an average score of 63.35 on the Open LLM Leaderboard, with notable scores in HellaSwag and TruthfulQA. It is designed for general-purpose language tasks, leveraging its merged architecture for broad applicability within its 4096-token context window.
Loading preview...
Blur-7b-v1.22: A Merged 7B Language Model
Blur-7b-v1.22 is a 7 billion parameter language model developed by liminerity, created through a strategic merge of three distinct models: s3nh/Sonya-Panda-7B-slerp, argilla/distilabeled-Marcoro14-7B-slerp, and Weyaxi/MetaMath-OpenHermes-2.5-neural-chat-v3-3-Slerp. This merge was executed using the TIES method via LazyMergekit, building upon the liminerity/Blur-7b-v1.21 base model.
Key Capabilities & Performance
This model exhibits a well-rounded performance profile, as indicated by its evaluation on the Open LLM Leaderboard. It achieved an average score of 63.35, demonstrating proficiency across several key metrics:
- HellaSwag (10-Shot): 82.00
- TruthfulQA (0-shot): 68.01
- AI2 Reasoning Challenge (25-Shot): 62.29
- MMLU (5-Shot): 58.03
- Winogrande (5-shot): 78.61
- GSM8k (5-shot): 31.16
These scores suggest a model capable of handling a variety of tasks, from common sense reasoning and factual recall to language understanding and basic mathematical problem-solving, within its 4096-token context window.
Good For
- General-purpose text generation and understanding: Its balanced benchmark performance makes it suitable for a wide array of language tasks.
- Applications requiring a blend of reasoning and factual knowledge: The merge incorporates models with strengths in different areas, contributing to its versatility.
- Developers seeking a 7B model with a unique merged architecture: Offers an alternative to single-base models, potentially providing a distinct blend of capabilities.