Microsoft Hacked World of Mathematics: Orca-Math Shines

In the world of education and research, STEM (science, technology, engineering and mathematics) has occurred a significant event. Microsoft introduced a new model of artificial intelligence called Orca-Math , which significantly exceeds existing models in solving mathematical problems, while having 10 times smaller size. This discovery became possible thanks to cooperation with the French startup Mistral and its Mistral 7B model, which is a variant of the LLAMA 2 model from META.

Orca-Math demonstrates the outstanding ability to solve “mathematical problems in verbal form”, while maintaining a small size for training and output. It is noted that the new model exceeds the majority of existing AI models with 7-70 billion parameters, giving way to only Gemini Ultra from Google and GPT-4 from OpenAi at the GSM8K benchmark, which consists of 8500 different mathematical tasks.

The creation of Orca-Math was made possible thanks to the development of a new set of 200,000 mathematical tasks using “specialized agents”, including students and teachers who corrected their answers. These tasks were created by analyzing and processing 36,227 examples from open sources, followed by training in the new version of the Mistral 7B model, which became Orca-Math.

The Orca team also used the Kahneman-Tversky optimization method (KTO method), developed by the startup of Contextual Ai, to improve the accuracy of the model’s answers to mathematical questions. This method, inspired by the work of economists Kaneman and Tverski on decision -making, allows you to improve the alignment of the model without the need for preferences such as “output B is better than exit X”

Microsoft placed a synthetic set of 200,000 mathematical tasks generated by AI on the Huging Face platform under the MIT license, providing everyone to explore, build and include with it, including commercial use.

This discovery reflects the tendency to use machine-generated data to increase the intellectual abilities and capabilities of large language models (LLM), thereby reducing fears regarding the “model collapse”. With each update, the Orca family from Microsoft continues to grow and becomes smarter, offering new, smaller in size or similar models that can solve more and more complex problems.

/Reports, release notes, official announcements.