Baichuan AI x Ascend AI | Baichuan 2 Open Sourced in the MindSpore Community

[Beijing, September 6, 2023] Baichuan AI held a conference in Beijing and officially released the Baichuan2 open-source foundation model, supported by the Ascend AI software and hardware platform. The Baichuan2-7B model is now freely available in the MindSpore community.

At the launch event, the company announced the official open-source of fine-tuned models Baichuan2-7B, Baichuan2-13B, Baichuan2-13B-Chat, and a 4-bit quantitative version. These models, offering various services at no cost, cater to both academic and commercial markets and are ready for commercial deployment.

(Launch event of the Baichuan2 model)

Link to the MindSpore open-source repository:

https://gitee.com/mindspore/mindformers/blob/dev/research/baichuan2/baichuan2.md

Link to the foundation model platform in the MindSpore open-source community:

https://xihe.mindspore.cn/modelzoo/baichuan2_7b_chat

Excellent Performance Surpassing LLaMA 2

Baichuan2-7B-Base and Baichuan2-13B-Base, trained on 2.6 TB of high-quality multilingual data, have notably enhanced their proficiency in areas like mathematics, coding, security, logical reasoning, and semantic understanding. Simultaneously, they maintain the optimal features of prior open-source models, including content generation, smooth multi-turn dialogues, and easy deployment. When contrasted with the preceding 13B model, Baichuan2-13B-Base exhibits enhancements in various capabilities: a 49% improvement in mathematics, 46% in coding, 37% in security, 25% in logical reasoning, and 15% in semantic understanding.

The two open-source models demonstrate superior performance in key evaluation rankings. They outperform LLaMA 2 in authoritative benchmarks like MMLU, CMMLU, and GSM8K. Even when compared to other models with an equivalent number of parameters, they surpass the performance of other competing models.

Notably, Baichuan2-7B, with seven billion parameters, performs on par with LLaMA 2, which has 13 billion parameters, in mainstream English tasks according to multiple authoritative benchmarks such as MMLU.

(Benchmark scores of models with 7 billion parameters)

(Benchmark scores of models with 13 billion parameters)

Baichuan2-7B and Baichuan2-13B not only are completely open to academic research, but also can be put into commercial use for free after developers apply for official commercial licenses by email.

Baichuan2 Foundation Model

Baichuan2, both open-source and commercially available, is a series of pre-trained large language models developed by Baichuan AI. The series includes models with 7 billion, 13 billion, and 53 billion parameters. From its establishment, Baichuan AI has been committed to fostering the growth of China's foundational model industry via open sourcing. The two open-sourced Baichuan2 models have garnered favorable feedback from both upstream and downstream businesses. Numerous renowned companies, including Huawei, attended the launch event and entered into collaborations with Baichuan AI.

Ascend AI

Ascend AI is an AI computing industry built based on the Ascend AI software and hardware platform. The platform includes Atlas series hardware and partner-branded hardware, heterogeneous computing architecture CANN, all-scenario AI framework MindSpore, Ascend application enablement MindX, one-stop development platform ModelArts, and unified toolchain MindStudio.