The editor of Downcodes learned that the Mistral AI team released a 7B mathematical model called MathΣtral, which has a 32k context window, can handle longer and more complex mathematical problems, and is open source under the Apache2.0 license. MathΣtral achieved 56.6% on the MATH benchmark and 63.47% on the MMLU benchmark. Through majority voting and reward models, the scores were as high as 68.37% and 74.59%. This is not only a tribute to the 2311th anniversary of Archimedes, but also a major breakthrough in the fields of mathematical reasoning and scientific discovery, demonstrating Mistral AI's efforts in supporting academic projects.
The Mistral AI team contributes MathΣtral to the scientific community, hoping to strengthen research on advanced mathematical problems that require complex, multi-step logical reasoning. The model's professional expertise in the STEM field has achieved the same category of advanced reasoning capabilities in various industry standard benchmark tests. In particular, it achieved 56.6% on the MATH benchmark and 63.47% on the MMLU benchmark. What's most striking about MathΣtral is its reasoning capabilities. This model demonstrates that significantly better results can be achieved with more inference time computations. In the MATH benchmark, MathΣtral7B achieved a score of 68.37% through majority voting, and an even higher score of 74.59% among 64 candidates through a powerful reward model. This move by the Mistral AI team is part of the company’s broader efforts to support academic projects. The release of MathΣtral was produced in the context of cooperation with Project Numina and reflects Mistral AI's emphasis on and support for academic research. MathΣtral is a guided model that can be used or fine-tuned according to Mistral AI's documentation. Model weights are hosted on HuggingFace, and now users can try MathΣtral using misstral-inference and adapt it to meet specific needs using misstral-finetune. Mistral AI's MathΣtral model is not only a leap in technology, but also a profound contribution to research in the fields of mathematics and science. With the continuous development of AI technology, we have reason to believe that MathΣtral will bring more possibilities and breakthroughs to mathematical reasoning and scientific discovery.
Official website address: https://mistral.ai/news/mathstral/
The open source and powerful reasoning capabilities of the MathΣtral model have brought new tools and possibilities to mathematics and scientific research, which are worthy of attention and anticipation. The editor of Downcodes will continue to pay attention to new developments in the field of AI and bring more exciting content to readers.