Baichuan Intelligent released Baichuan 2 open source model, which claims to be "leading LLaMA 2 in both arts and science".

CTOnews.com Sept. 6 news, Baichuan Intelligence today held a large model conference with the theme of "Baichuan Huihai, Open Source win-win". Officially open source fine-tuned Baichuan2-7B, Baichuan2-13B, Baichuan2-13B-Chat and its 4bit quantitative version, and are free and commercially available.

Baichuan Intelligent also opened up Check Point for model training this time, and announced that it would release a Baichuan 2 technical report detailing the training details of Baichuan 2. CTOnews.com is attached to Baichuan 2 download address: https://github.com/baichuan-inc/Baichuan2

According to the official introduction of Baichuan Intelligence, both Baichuan2-7B-Base and Baichuan2-13B-Base are trained based on 2.6 trillion high-quality multilingual data, and the two models have significantly improved their abilities in mathematics, code, security, logical reasoning and semantic understanding.

Compared with the previous generation 13B model, Baichuan2-13B-Base increases math ability by 49%, code ability by 46%, security ability by 37%, logical reasoning ability by 25%, and semantic understanding ability by 15%.

The official account of "Baichuan Intelligence", which is the source of ▲, is also in the benchmark tests of MMLU, CMMLU, GSM8K, etc., the two models of Baichuan Intelligence have higher scores than LLaMA 2, and have a greater advantage in the "Chinese universal" part.

Benchmark score of ▲ 7B Parametric Model

The Benchmark scores of ▲ 13B parameter model Baichuan2-7B and Baichuan2-13B are not only completely open to academic research, but also developers can do business free of charge after applying for an official commercial license by email. Tencent Cloud, Aliyun, Volcano Ark, Huawei, MediaTek and other companies all attended the conference and reached cooperation with Baichuan Intelligence.