MathΣtral
7B scale mathematical reasoning and scientific discovery model
- Specially designed for mathematical reasoning and scientific discovery, suitable for solving complex mathematical problems.
- Having a 32k contextual window, capable of processing a large amount of information.
- High accuracy rates of 56.6% and 63.47% were achieved in the MATH and MMLU benchmark tests, respectively.
- Through majority voting and strong reward models, Mathstral 7B can further improve to 68.37% and 74.59% in the MATH benchmark test.
- As a guiding model, it can be used or fine tuned through our documentation.
- The weight is hosted on the HuggingFace platform, which can be tested through wrong reference and customized through wrong finetune.
Product Details
Math ∑ tral is a 7B scale AI model designed for mathematical reasoning and scientific discovery, with a 32k context window, released under the Apache 2.0 license. It demonstrates outstanding performance in advanced mathematical problems involving multi-step complex logical reasoning, and is a contribution of the Mistral AI team to the scientific community, aimed at strengthening support for academic projects. Math ∑ tral has professional expertise in the STEM field, and its reasoning ability has reached the forefront level of industry standard benchmarks among similar scale models.
