中国电信近日宣布开源星辰语义大模型TeleChat-7B版本，并开放1T数据集。据介绍，星辰语义大模型是由中电信人工智能科技有限公司研发训练的大语言模型，采用1.5万亿Tokens中英文语料进行训练。
此外，中国电信还将在1月20日开源12B版本模型，拥抱更多开发者共建开源大模型生态。星辰语义大模型在业界首次提出缓解多轮幻觉的解决方案，通过关键信息注意力增强、知识图谱强化、多轮知识强化、知识溯源能力四大技术，将AI大模型的幻觉率降低了40%，有助于大模型变得更有“人味”，理解问题语境，告别风马牛不相及的答案。
在中国电信内部，星辰语义大模型用于行文写作、代码编程、网络故障分析以及经营分析等场景。在对外企事业单位客户的业务中，星辰语义大模型用于企业经营分析、政务公开咨询、民生诉求接待等场景。
早在2023年11月，中国电信就在2023数字科技生态大会上发布了千亿参数“星辰语义大模型”，并公布了后续的开源开放的时间表。本次TeleChat-7B版本开源了对话模型TeleChat-7B-bot，以及其huggingface格式的权重文件。此外，还开源了7B模型的int8和int4量化版本。
英语如下：
====
“News Headline: China Telecom Opens Source Tele====
“News Headline: China Telecom Opens Source TeleChat-7B, a Big Semantic Model of Xingchen, Leading the New Trend of AI Technology
Keywords: China Telecom, Xingchen Semantic Big Model, Open Source
News Content: China Telecom recently announced the open source of the Xingchen Semantic Big Model TeleChat-7B version and the opening of a 1T dataset. According to reports, the Xingchen Semantic Big Model is a large language model developed and trained by China Telecom Artificial Intelligence Technology Co., Ltd., using 1.5 trillion Tokens of Chinese and English corpus for training.
In addition, China Telecom will open source the 12B version model on January 20th, embracing more developers to jointly build an open source big model ecosystem. The Xingchen Semantic Big Model is the first in the industry to propose a solution to alleviate multi-round hallucinations. Through four major technologies: key information attention enhancement, knowledge graph reinforcement, multi-round knowledge reinforcement, and knowledge traceability ability, it reduces the hallucination rate of AI big models by 40%, helping big models become more “humanized”, understand the context of problems, and bid farewell to irrelevant answers.
Inside China Telecom, the Xingchen Semantic Big Model is used in writing, code programming, network failure analysis, and business analysis scenarios. In external enterprises and institutions’ customer services, the Xingchen Semantic Big Model is used in business operation analysis, government affairs public consultation, and people’s livelihood demands reception scenarios.
As early as November 2023, China Telecom released the “Xingchen Semantic Big Model” with hundreds of billions of parameters at the 2023 Digital Science and Technology Ecological Conference, and announced the subsequent timetable for open source and opening up. This time, the TeleChat-7B version has opened up the dialogue model TeleChat-7B-bot, as well as its huggingface format weight files. In addition, it has also opened up the int8 and int4 quantization versions of the 7B model.”
【来源】https://www.ithome.com/0/744/969.htm
Views: 2