MathΣtral - Mistral AI

MathΣtral by Mistral AI is your pathway to advanced mathematical reasoning and scientific discovery. In tribute to the legendary Archimedes, they celebrate his 2311th anniversary by launching the Mathstral model, a 7B model with exceptional capabilities. Boasting a 32k context window, this model is designed to tackle complex math problems, making it a boon for the science community and academic endeavors. Releasing under the Apache 2.0 license, Mathstral is a state-of-the-art AI available for public use. With its fine-tuning features on la Plateforme, users can customize Mathstral for various STEM applications, reflecting the contemporary commitment to accessible and powerful AI tools.

Key Features

• High Context Window: The Mathstral model includes an extensive 32k context window, facilitating in-depth analysis and reasoning for mathematical problems.

• STEM Focus: Designed specifically for STEM subjects, Mathstral demonstrates superior reasoning capabilities, standing out in its category.

• Benchmark Performance: Achieves top-notch performance on industry-standard benchmarks, such as 56.6% on MATH and 63.47% on MMLU.

• Enhanced Inference Computation: With majority voting and a strong reward model, Mathstral significantly improves its results, boasting up to 74.59% accuracy.

• Open Access and Customization: Mathstral's weights are hosted on HuggingFace, allowing developers to integrate, use, and fine-tune the model as needed.

FAQ

What is Mathstral?

Mathstral is a specific 7B model designed for advanced mathematics and scientific exploration.

What benchmarks does Mathstral perform on?

Mathstral achieves a performance of 56.6% on MATH and 63.47% on MMLU benchmarks.

Can Mathstral's performance be improved?

Mathstral can be fine-tuned and further adapted through majority voting and reward models, which can improve its performance up to 74.59% accuracy.

Is Mathstral available for public use?

Yes, Mathstral is available for public use and can be accessed for integration and customization through HuggingFace.

Who contributed to the development of Mathstral?

Mathstral was produced in collaboration with Project Numina as part of Mistral AI's efforts to support scientific and academic projects.