DatawhaleShare
Mathematical large model: MathGPT, source: Heart of the Machine
TAL launched MathGPT, a 100-billion-level large-scale model in the field of mathematics, to do a good job in the basic work of mathematics in the AI era.
The domestic large-scale model market has ushered in a new "player", this time it is a large-scale model dedicated to mathematics.
On August 24, the heart of the machine learned that in the live broadcast event of the 20th anniversary of TAL, CTO Tian Mi announced that MathGPT, a 100-billion-level large-scale model in the mathematics field developed by TAL, started internal testing. From now on, users can apply for a free trial experience through the official website ( www.mathgpt.com ) to register an account.
In May of this year, TAL announced that it is developing a self-developed large mathematical model, named MathGPT. MathGPT is a large-scale model in the vertical field of mathematics with problem-solving and lecture algorithms as the core for mathematics enthusiasts and scientific research institutions around the world. It is also the first large-scale model specially built for mathematics in China.
It is also very simple to use. When users use MathGPT, they can upload math questions in text or pictures, and then they can get dialogue-style answer feedback. They can also use the "random question" button to randomly generate math questions and give answers by the system.
Currently, MathGPT supports PC and mobile experiences in Chinese and English versions.
Leading mathematical problem-solving skills
MathGPT brings together TAL's years of education, teaching and research data accumulation, focusing on the field of mathematics. The training, reasoning, and deployment framework of hundreds of billions of large models endows the model with powerful capabilities. Through high-quality educational data, multi-task continuous training and supervised fine-tuning such as topic calculation, explanation, and question-and-answer are realized, showing excellent performance. In addition, with the help of human feedback alignment, the comprehensive quality of the model will be further improved. MathGPT has obvious advantages in problem-solving accuracy, stability and user experience.
It is understood that MathGPT's mathematical computing ability has covered mathematics problems in elementary school, junior high school, and high school. However, Q&A interactions other than mathematics are not yet open.
MathGPT Technical Report
What is the specific effect? Among the test results of six public mathematics evaluation collections including CEval-Math, AGIEval-Math, APE5K, CMMLU-Math, Gaokao Mathematics and Math401, MathGPT achieved the highest scores in multiple tests. At the same time, MathGPT also performed well on the general test set of C-Eval's middle and high schools.
MathGPT's C-Eval list of junior and senior high school subjects
In addition, in terms of problem-solving stability and explanation friendliness, MathGPT conducts model training based on a large number of famous teachers' problem-solving process data, and the model's problem-solving steps are professional and clear.
Let's take a sequence question as an example. The answer given by MathGPT includes three parts: "analysis", "detailed explanation" and "finishing", which is more detailed than the rough explanation of the general large model. Among them, "analysis" provides the problem-solving ideas and thinking methods of the topic, helping users better understand the topic; Click to prompt, help users review and reflect on the intention of the question, and draw inferences from one instance.
For users, researching mathematical problems is not only about getting the answers themselves, but also about the problem-solving principles and thinking logic behind the answers. Compared with other general-purpose large models, MathGPT can solve problems with higher accuracy, and can also analyze the answers more clearly and explain them more clearly, and better meet the core needs of users to use AI products to solve mathematical problems.
At the same time as MathGPT was released, TAL also updated a representative and challenging math task evaluation set for global artificial intelligence experts and math enthusiasts to experience and evaluate. TAL hopes to make MathGPT play a greater role in the field of mathematics education, and is willing to share with the industry the experience and methods of developing hundreds of billions of large models based on large-scale, high-quality content, and make progress together with the industry.
TAL's Accumulation of AI
Driven by the wave of AI, many technology companies have announced the launch of general-purpose large language model products this year, but TAL has chosen another direction, not based on fine-tuning and interface calling of existing large language models, and not making general-purpose large language models , but in-depth research and development of large models in the vertical field of mathematics, and is committed to creating independent, stable, sustainable, and high-quality mathematical solutions.
The general large-scale model "emphasizes literature over theory", but there are obvious shortcomings in the solving, explanation, question and answer and recommendation of mathematical problems. On another level, on the road to general artificial intelligence, mathematical reasoning ability is very important, and many large companies around the world are doing research in this area.
"Talk Future has 20 years of accumulation in mathematics data and business, and has accumulated a large amount of educational data and the ability to continuously produce educational data, so it chooses to do this difficult but correct thing." Tian Mi said that Talent Future hopes to use With my years of accumulation in mathematics and AI, I will do a good job in the basic work of mathematics in the era of AI large models.
In fact, as early as 2017, TAL established the AI lab artificial intelligence laboratory. Based on the support of the smart education AI open innovation platform, TAL AI lab has won 16 championships and 6 runner-ups in various top academic conference competitions, and published nearly 100 high-level academic papers in international journals and conferences.
In 2019, the Ministry of Science and Technology announced that relying on TAL to build a national next-generation AI open innovation platform for smart education, TAL has become the first and only member of the AI "national team" in the education industry, and has years of in-depth research in the field of AI. Over the years, TAL has built a national education technology innovation platform with education-oriented artificial intelligence algorithm capabilities, application solutions, basic software and hardware systems, and open source and open services driven by the major needs of the education industry.
TAL is also actively participating in the promotion of the construction of a large-scale model standard system. As a core unit, it has successively participated in the large-scale model series national standards organized by the National Artificial Intelligence Standardization General Group, and the "Large-scale pre-training model technology and application evaluation" led by the China Academy of Information and Communications Technology. Methods" series of group standards, and the preparation of the "Education General Model" series of standards led by the Educational Informatization Technology Standards Committee of the Ministry of Education and the National Information Technology Standardization Technical Committee.
Recently, TAL is taking the leading role in compiling group standards for educational large models together with China Academy of Information and Communications Technology, Fudan University, iFLYTEK, Baidu and other industry-leading scientific research institutions, universities, and enterprises, comprehensively covering scenarios, application effects, and service reliability. Evaluate the capabilities of large educational models, and provide reference and guidance for the application of large educational models.
Using AI to achieve large-scale individualized teaching
With the rise of large language models, how to use AI technology to serve all walks of life is the focus of social attention. The education industry is one of the first industries to start to lay out the AI field, and what changes AI can bring to the education ecology has always attracted attention.
"AI brings the opportunity to redefine the education industry, and large-scale model technology makes it possible to realize large-scale individualized teaching." Tian Mi introduced that TAL has been exploring personalized learning for 20 years, from offline small classes to online learning. From large classes to AI classes, the form is constantly evolving, but the teaching content is always fixed, the interaction between students and teachers is less, and the granularity can only reach the topic level.
Tian Mi believes that the essence of large models is a more efficient way to learn knowledge from data and apply it. With the blessing of AI capabilities, the new learning method of "students' self-study + AI answering questions" has become widely possible. The threshold and cost for learners to obtain high-quality teaching content are reduced, and the degree of personalization and refinement of the obtained teaching content continues to increase. It can realize AI teaching and Q&A tutoring for thousands of people, and every student can get the most suitable learning for them content.
Based on MathGPT, TAL will continue to explore learning methods in an AI environment to better serve learners and math lovers around the world, share its experience with the industry in a timely manner, and help positive changes in educational technology through AI technology.
With the smooth progress of the internal test, MathGPT's problem-solving ability will continue to improve, and product-level applications based on MathGPT are also being accelerated and will be released in the near future.
Dry goods learning, like three times ↓