Smaug-2-72B: Enhanced Reasoning and Coding

Smaug-2-72B is a 72.3 billion parameter language model developed by abacusai, building upon the Qwen1.5-72B-Chat architecture. This iteration of Smaug has undergone specialized fine-tuning to excel in complex reasoning and coding challenges, making it a powerful tool for developers and researchers.

Key Capabilities & Performance

• Superior Reasoning: Smaug-2-72B shows enhanced reasoning abilities, outperforming its base model, Qwen1.5-72B-Chat, on the MT-Bench benchmark with an average score of 8.53 compared to 8.34.
• Strong Coding Proficiency: The model demonstrates significant improvements in code generation, achieving a pass@1 score of 66.5% on HumanEval, which is approximately 10% higher than Qwen1.5-72B-Chat's 56.7%.
• Fine-tuned from Qwen1.5-72B-Chat: Leveraging the robust foundation of Qwen1.5-72B-Chat, Smaug-2-72B integrates new techniques and data for its specialized performance.

When to Use This Model

• Complex Reasoning Tasks: Ideal for applications requiring advanced logical inference, problem-solving, and understanding intricate relationships.
• Code Generation and Assistance: Highly suitable for developers needing assistance with writing, debugging, or understanding code, particularly in scenarios where high accuracy is critical.
• Benchmarking and Research: Offers a strong baseline for further research into fine-tuning techniques for reasoning and coding, building on the insights from the previous Smaug paper.