Alibaba'sThousand Questions on Tongyi"The team has made another big news! They just released the Qwen2Math Demo, thisMathematical modelIt's just a little monster.GPT-4All were trampled under its feet.
This model can not only handle math problems entered in text, but also understand formulas in pictures and screenshots. Imagine that you take a picture of a math equation and it can give you the answer. It is simply a magic tool for solving math problems in math class! (Of course, we do not encourage cheating)
Qwen2-Math has launched three versions: 72B, 7B and 1.5B. The 72B version is simply a math genius. It scored 7 points more than GPT-4 on the MATH dataset, an improvement of 9.6%. This is like you scored 145 points in the college entrance examination math, while the top student next to you only scored 132 points.
What's more impressive is that the 7B version uses less than one-tenth of the number of parameters, but it has surpassed the 72B open source mathematical model NuminaMath. You should know that NuminaMath is the model that won the first AIMO in the world, and the award was personally presented by Terence Tao, the "top boss" in the mathematics community.
Lin Junyang, a senior algorithm expert at Alibaba, excitedly announced that they had turned the Qwen2 model into a math expert. How did they do it? They used a special "mathematical brain tonic" - a carefully designed math corpus. This "brain tonic" contains a large number of high-quality math online texts, books, codes, exam questions, and even math questions "compiled" by the Qwen2 model itself.
The result? In classic math test sets such as GSM8K and MATH, Qwen2-Math-72B left 405B's Llama-3.1 behind. These test sets are no joke, they contain algebra, geometry, probability, number theory and other math problems.
Not only that, Qwen2-Math also challenged the Chinese dataset CMATH and college entrance examination questions. On the Chinese dataset, even the 1.5B version can beat the 70B Llama3.1. Moreover, no matter which version, the results are significantly improved compared with the Qwen2 basic model of the same scale.
It seems that "Tongyi Qianwen" has really found a math genius this time! Can we ask it questions when we do math problems in the future? But remember, it is just a tool. Don't be fooled by its intelligence. You still need to practice your math skills well!
Online experience address: https://huggingface.co/spaces/Qwen/Qwen2-Math-Demo