nfaheem/SOLAR-10.7B-Instruct-ties
nfaheem/SOLAR-10.7B-Instruct-ties is a 10.7 billion parameter instruction-tuned language model created by nfaheem, formed by merging kodonho/Solar-OrcaDPO-Solar-Instruct-SLERP and VAGOsolutions/SauerkrautLM-SOLAR-Instruct using the TIES merging method. This model leverages a 4096-token context length and achieves an average benchmark score of 74.24 across various tasks, including 66.34 on MMLU and 64.06 on GSM8K. It is designed for general instruction-following tasks, demonstrating strong performance in reasoning and language understanding.
Loading preview...
Model Overview
nfaheem/SOLAR-10.7B-Instruct-ties is a 10.7 billion parameter instruction-tuned language model. It was created by nfaheem through a merge of two existing models, kodonho/Solar-OrcaDPO-Solar-Instruct-SLERP and VAGOsolutions/SauerkrautLM-SOLAR-Instruct, utilizing the TIES merging method. The base model for this merge was upstage/SOLAR-10.7B-Instruct-v1.0.
Key Capabilities & Performance
This model is designed for general instruction-following and demonstrates solid performance across a range of benchmarks. It features a context length of 4096 tokens.
- Average Score: Achieves an average score of 74.24 across evaluated benchmarks.
- Reasoning: Scores 70.9 on ARC and 64.06 on GSM8K, indicating capabilities in reasoning and mathematical problem-solving.
- Language Understanding: Demonstrates strong performance with 88.58 on HellaSwag and 83.5 on Winogrande.
- Knowledge & Truthfulness: Scores 66.34 on MMLU and 71.88 on TruthfulQA.
Use Cases
SOLAR-10.7B-Instruct-ties is suitable for applications requiring a capable instruction-following model, particularly where a balance of performance and model size is desired. Its benchmark results suggest it can be effectively used for tasks involving general knowledge, common sense reasoning, and language generation.