Shulou(Shulou.com)11/24 Report--

Thanks to CTOnews.com netizens Wu Yanzu in South China for the delivery of clues! CTOnews.com September 25 news, Baichuan Intelligent today announced the launch of the Baichuan2-53B closed-source large model, a comprehensive upgrade of the capabilities of Baichuan-53B.

According to reports, Baichuan2-53B not only significantly improves the ability of mathematical and logical reasoning, but also greatly reduces model hallucinations through high-quality data systems and search enhancements, and is known as "the largest model with the lowest hallucination problem in China."

Baichuan Intelligent also opened the Baichuan2-53B API interface this time, officially entered the To B field, and started the commercialization process.

CTOnews.com query found that Baichuan2-53B is the sixth large model released by Baichuan Intelligence. Since its establishment on April 10, Baichuan Intelligence has launched an average of one large model every 28 days.

Baichuan Intelligence said that on the basis of Baichuan-53B, Baichuan2-53B not only focuses on strengthening the capabilities of the above two dimensions, but also comprehensively upgrades the overall capabilities. Among them, logical reasoning ability increased by 100%, mathematics ability increased by 31%, language comprehension ability increased by 29%, text creation increased by 18%, and knowledge question and answer increased by 9%.

Baichuan2-53B is also optimized in hallucination processing. In terms of high-quality data construction, Baichuan2-53B claims to have "created an original data quality system". Classify the data with low quality and high quality as the standard to ensure that the model always uses high quality data for pre-training.

In the aspect of information acquisition, Baichuan2-53B upgrades a number of modules, including key components such as instruction intention understanding, intelligent search and result enhancement. This comprehensive system through in-depth understanding of user instructions, accurately drive query word search, and finally combined with large language model technology to optimize the reliability of model results, achieve more accurate and intelligent model answer results, and reduce model hallucinations.

The results evaluated by Baichuan Intelligence using FacTool show that the comprehensive score of Baichuan2-53B is 140.5, which only ranks behind GPT-4 in the mainstream basic model.

CTOnews.com Note: FacTool is a general framework proposed by researchers from Shanghai Jiaotong University, Carnegie Mellon University, City University of Hong Kong, Meta and other institutions, which can check the factual accuracy of the content generated by the large model (but also check the factual accuracy of general content).

In addition to the overall capability upgrade, Baichuan2-53B also opens the API interface, which allows enterprises and developers to integrate Baichuan2-53B into their applications and services through API. CTOnews.com query found that the API invocation service of Baichuan2-53B is charged according to time periods:

From 8:00 to 24:00 daily, at the rate of 0.02 yuan per thousand tokens

From 00:00 to 8:00 daily, it is charged at the rate of 0.01 yuan per thousand tokens.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.