PrimeIntellect/INTELLECT-MATH-SFT
INTELLECT-MATH-SFT is a 7.6 billion parameter language model developed by PrimeIntellect, specifically optimized for advanced mathematical reasoning tasks. This model serves as the supervised fine-tuning (SFT) stage for the INTELLECT-MATH series, leveraging verified QwQ outputs to imitate strong teacher model reasoning. It demonstrates strong performance across various mathematical benchmarks, including MATH-500, OLYMPIADBENCH, and AIME 2024, making it suitable for complex problem-solving in mathematics.
Loading preview...
INTELLECT-MATH-SFT: Foundation for Mathematical Reasoning
INTELLECT-MATH-SFT is the supervised fine-tuning (SFT) stage of the 7.6 billion parameter INTELLECT-MATH model developed by PrimeIntellect. This model is specifically designed and optimized for advanced mathematical reasoning. Its training involved fine-tuning on a high-quality synthetic dataset derived from verified QwQ outputs, which encourages the model to emulate the reasoning processes of a powerful teacher model.
Key Capabilities and Performance
INTELLECT-MATH-SFT provides a robust foundation for mathematical problem-solving, demonstrating strong performance across a range of challenging benchmarks. Its SFT data quality is crucial, as it significantly impacts the subsequent reinforcement learning (RL) stage, leading to faster training and improved final performance. While INTELLECT-MATH-SFT itself is the initial stage, its scores highlight its strong base in mathematical understanding:
- MATH-500: Achieves 72.8
- OLYMPIADBENCH: Scores 39.1
- AIME 2024: Reaches 16.6
- AMC: Scores 45.8
- MINERVA MATH: Achieves 33.8
These results indicate its proficiency in handling diverse mathematical challenges, from standard problems to competitive-level tasks.
Use Cases
INTELLECT-MATH-SFT is ideal for applications requiring:
- Mathematical problem-solving: Excelling in areas like algebra, geometry, and number theory.
- Educational tools: Assisting in generating explanations or solutions for math problems.
- Research in AI for mathematics: Serving as a strong baseline or component in more complex mathematical AI systems.