nvidia/NFT-7B
NFT-7B is a 7.6 billion parameter math reasoning model developed by NVIDIA, Tsinghua University, and Stanford University. Fine-tuned from Qwen2.5-Math-7B using the Negative-aware Fine-Tuning (NFT) algorithm, it excels at mathematical problem-solving by learning from both correct and incorrect answers. This model is specifically optimized for competition-level mathematics and general mathematical reasoning, demonstrating significant performance improvements over its base model.
Loading preview...
Model Overview
NFT-7B is a 7.6 billion parameter model developed by NVIDIA, Tsinghua University, and Stanford University, specifically designed for advanced mathematical reasoning. It is fine-tuned from Qwen2.5-Math-7B using the innovative Negative-aware Fine-Tuning (NFT) algorithm. This supervised learning approach allows the model to learn from its failures by constructing an implicit negative policy, achieving performance comparable to leading reinforcement learning methods without external teachers.
Key Capabilities
- Advanced Mathematical Reasoning: Excels in competition-level mathematics (AIME, AMC, Olympiad) and general mathematical reasoning (MATH500, Minerva Math).
- Self-Reflective Improvement: Leverages negative feedback to continuously improve its problem-solving abilities.
- Step-by-Step Solution Generation: Capable of generating detailed, step-by-step mathematical solutions.
- LaTeX Support: Handles mathematical expressions using LaTeX notation for both input and output.
Performance Highlights
NFT-7B demonstrates substantial performance gains over its base model, Qwen2.5-Math-7B, across various benchmarks. It shows an average improvement of +20.1% across six mathematical reasoning datasets, including a +42.7% increase on AMC23 and +18.7% on AIME24.
Good for
- Researchers and developers working on advanced mathematical AI.
- Applications requiring robust, step-by-step mathematical problem-solving.
- Benchmarking and exploring novel supervised learning techniques for self-improvement in LLMs.
Limitations
- Domain Specificity: Primarily designed for mathematical tasks; not recommended for general conversation.
- Calculation Errors: May still exhibit arithmetic errors in highly complex calculations.
- Context Understanding: Struggles with problems requiring real-world context outside of mathematics.