超大规模多模态预训练模型M6的关键技术突破及产业应用
发布时间:2022-04-08 作者:林俊旸,周畅,杨红霞 阅读量:
超大规模多模态预训练模型M6的关键技术及产业应用
林俊旸, 周畅, 杨红霞
(阿里巴巴达摩院,中国 杭州 311100)
摘要:阿里巴巴达摩院研发了超大规模中文多模态预训练模型M6,并陆续推出了百亿、千亿、万亿和十万亿参数规模的预训练模型,实现了高效低碳的预训练,推动超大规模预训练模型的产业化应用。同时,推出了M6服务化平台,帮助广大用户快速使用大模型。未来,大模型在产业领域的应用将更加丰富。
关键词:多模态预训练;大规模预训练;图像生成;文本生成
Key Technologies and Applications of Extremely Large-Scale Multimodal Pre-Trained Model M6
LIN Junyang, ZHOU Chang,YANG Hongxia
(Alibaba DAMO Academy, Hangzhou 311100, China)
Abstract: The extremely large-scale Chinese multimodal pre-trained model M6 is proposed by Alibaba DAMO Academy, and the 10 B, 100 B, 1 T, and 10 T versions of M6 are released. M6 has been trained efficiently with low carbon emission, and it has been deployed in multiple scenarios, which leads to the creation of new products as well as performance improvement. Also, to provide better services, the easy-to-use M6 platform for users to leverage large-scale pre-trained models is released by DAMO Academy.
Keywords: multimodal pre-training; large-scale pre-training; image generation; text generation
本期相关文章