A group of Huawei researchers developed a system that trained a language model—PanGu-Σ under the framework of MindSpore 5 on a cluster of Ascend 910 AI processors with 329 billion tokens over 100 days and launched it towards the second half of March.

PanGu-Σ’s built-in parameters are expanded using Random Routed Experts and the Transformer decoder architecture from PanGu-α‘s Random Routed Experts.

It is simple to extract sub-models using RRE design from the PanGu-Σ for a variety of downstream applications, including conversation, translation, code production, and interpreting natural language in general.

According to the research paper—in total, the training throughout is 6.3 times faster than it was for the model with the MoE architecture but the same hyper-parameters. The sub-modal of PanGu-Σ in the Chinese domain significantly outperforms the previous SOTA models, including PanGu-α- with 13 billion parameters and ERNIE 3.0 Titan with 260 billion parameters over 16 downstream tasks in six categories in the zero-shot setting without any multitask finetuning or instruction tuning. It uses 329 billion tokens in more than 40 natural and programming languages.

Huawei gathere datasets in 40 domains, with a significant amount of data in four key domains: Chinese, English, Bilingual (Chinese and English), and code, to further illustrate the PanGu-Σ’s ability model’s to learn effectively and independently from many domains.

The research paper asserts that PanGu-Σ has successfully produced state-of-the-art results in a variety of downstream tasks like few-shot NLU, open-domain discussion, question answering, machine translation, and code creation by expanding and continuously training from PanGu-α using 329B tokens.