Upstage's CEO Kim Seong-hoon (Courtesy of Upstage) South Korean artificial intelligence tech startup Upstage said on Monday that its math-specific large language model (LLM), jointly developed with local startup Masspresso and telecom leader KT Corp., has outperformed Microsoft Corp.’s ToRA in two global math benchmark tests.
Upstage’s MathGPT achieved 0.488 out of a full score of 1 in the latest MATH benchmark test for LLMs having 13 billion parameters or less. The test is based on a dataset of 12,500 challenging math problems.
The Korean model outperformed OpenAI's LLM GPT-4, which scored 0.425, chatbot ChatGPT's 0.355 and ToRA's 0.481, Upstage said.
In the GSM8K benchmark, or Grade School Math 8K, MathGPT topped the LLM list. The Korean AI scored 0.782, beating ToRA’s 0.758. The benchmark is based on a dataset of 8,500 high quality, linguistically diverse grade school math word problems.
Math has been a difficult field in which to apply LLMs due to the need for logical reasoning and abstract thinking.
Masspresso, which collects around 10 million data on math problems and explanations per day, has provided Upstage with the dataset.
KT operates Korea’s largest graphics processing unit (GPU) farm, a set of servers that allocate resources to quickly perform calculations, to accelerate the two startups’ math-specific LLM development.
Upstage will lead the innovation of generative AI in math and other domains with its global top LLM tech, said Chief Executive Kim Seong-hoon.
AI in the global edtech industry, which has been at the level of Google search, will be upgraded with MathGPT, said a Quanda official.
We use cookies to provide the best user experience. By continuing to browse this website, you will be considered to accept cookies. Please review our Privacy Policy to learn our cookie policy.