Sakura-SOLAR-Instruct Overview

Sakura-SOLAR-Instruct is an instruction-tuned language model developed by Kyujin Han (kyujinpy) as part of an LLM research consortium with Media Group Saramgwasup and Marker. This model was constructed using the Mergekit method, indicating a focus on combining existing models to achieve enhanced performance. Detailed information regarding its development, including training and code, is available in the ⭐Sakura-SOLAR repository.

Key Capabilities

• Strong General Instruction Following: Achieved an average score of 74.40 on the Open LLM Leaderboard, securing the top rank on December 27, 2023.
• Reasoning and Common Sense: Demonstrates solid performance in reasoning tasks with scores like 70.99 on ARC and 83.66 on Winogrande.
• Knowledge and Factuality: Scored 66.33 on MMLU and 71.79 on TruthfulQA, indicating a good grasp of general knowledge and factual accuracy.
• Mathematical Reasoning: Achieved 65.20 on GSM8K, suggesting capabilities in mathematical problem-solving.

Good for

• Applications requiring a robust instruction-following model with strong general reasoning abilities.
• Tasks benefiting from a model with competitive benchmark performance across a range of academic and common sense evaluations.
• Developers looking for a model with a transparent development process, as training and code details are publicly shared.