AceMath-72B-Instruct: Advanced Mathematical Reasoning

AceMath-72B-Instruct is a 72.7 billion parameter model from NVIDIA's AceMath family, specifically engineered for mathematical reasoning. It is built upon the Qwen2.5-Math-72B-Base model and undergoes a multi-stage supervised fine-tuning (SFT) process, first with general-purpose SFT data, then with math-specific SFT data.

Key Capabilities

• Mathematical Problem Solving: Excels at solving English mathematical problems using Chain-of-Thought (CoT) reasoning.
• High Performance: Outperforms Qwen2.5-Math-72B-Instruct (71.8 vs. 68.2), GPT-4o (67.4), and Claude 3.5 Sonnet (65.6) on various math reasoning benchmarks.
• Reward Model Integration: Designed to work in conjunction with AceMath-72B-RM, an outcome reward model that evaluates and scores mathematical solutions.
• Extensive Context: Features a context length of 131072 tokens.

When to Use This Model

• Dedicated Math Tasks: This model is specifically recommended for applications requiring robust mathematical problem-solving capabilities.
• Research in Mathematical AI: The release of all training data supports further research in this field.

For general-purpose tasks including code and general knowledge, NVIDIA also offers the AceInstruct series of models. AceMath models are intended for non-commercial use.