Shulou(Shulou.com)11/24 Report--

Recently, in the large-scale pre-training model standard compliance verification organized by China Institute of Information and Communication, the model development module of LinSeer, a large-scale private domain model built by Xinhua 3 Group, a subsidiary of Ziguang, was rated as 4 percent, reaching the advanced level in China. This evaluation carried out an omni-directional and multi-dimensional evaluation of LinSeer in five functional modules: data management, model training, model management, model deployment and development process integration, which annotates the leading strength of Xinhua 3 in the field of large models in the private domain, and will provide more powerful support for all trades and industries to enter the AIGC era.

Five capability modules demonstrate the strength of the model with hard core technology.

As the heat of AIGC continues to rise, the development process of AI large model engineering and universalization is significantly accelerated. In order to further promote the rapid development of large model, China Institute of Information, Communication and Research issued trusted AI large model standard system 2.0, which provides a detailed reference for scientific evaluation of large model technical capability and application efficiency. The model development capability domain project tested by Xinhua III aims to comprehensively evaluate the large model development capability of the research subject. In the evaluation of five types of indicators, LinSeer has shown excellent technical level.

● data management: focus on verifying the data processing and data versioning management capabilities of large models, including data cleaning, data tagging, data quality inspection, etc. Through efficient data set management and data processing means, and combined with Oasis platform data quality inspection, LinSeer realizes further data tagging and management. The results show that it has outstanding performance in content completeness and functional support for data cleaning, and can fully support text, picture, audio and video data tagging.

● model training: focus on verifying the ability of large models to support multiple training methods, visualization and resource optimization scheduling. Based on the overall architecture of MaaS (models and Services), Xinhua 3 can provide comprehensive large model training and fine-tuning services to generate customized and exclusive models. The results show that LinSeer fully supports multi-mode, pre-training tasks, natural language and programming language in the training mode, with an average incremental accuracy of 91.9% and a resource utilization rate of 90%.

● model management: focus on verifying the ability of large models to support model storage, version management, and log management. LinSeer enables large models to have memory and support accurate answers through vector storage and retrieval. The results show that it can fully support model storage capabilities such as file system management and image management, as well as version management capabilities such as metadata management, relationship maintenance, structure management and so on.

● model deployment: focus on verifying the ability of large models to support model fine-tuning, transformation, tailoring, and quantification. Through supporting a variety of fine-tuning algorithms, LinSeer flexibly meets the different data and model requirements of industry customers, and also provides a wide range of model multi-type conversion capabilities. In the face of the lightweight scenario of the model, it also supports the cutting and quantification ability of the model, and the reasoning delay speedup and memory footprint ratio are at an advanced level.

Integration of ● development process: focus on verifying the autonomy of large models in the development process. LinSeer, combined with Xinhua three full stack ICT infrastructure monitoring tools, can organically integrate all aspects of AI model development, provide a unified development platform and tools, help industry customers activate private domain models effectively, quickly build intelligent applications, and realize "model use freedom".

Create the overall solution of AIGC and quickly build a large private domain model of the industry

Xinhua third Group implements the AI in ALL strategy, integrates AI into the software and hardware product portfolio of Xinhua third line, and realizes full stack and full scene technology coverage. At the same time, it further puts forward the industry enabling strategy AI for ALL, that is, based on the in-depth understanding of the needs of the industry, integrate the capabilities of AI into end-to-end solutions, and at the same time help all industries upgrade intelligently with partner services.

In order to better promote AI application innovation and industrial landing, Xinhua Group further launched the overall AIGC solution, starting from the enabling platform, data platform and computing platform, to fully meet the business scene needs of users, and help customers quickly build a private domain model featuring industry focus, regional exclusive, data proprietary and value exclusive.

As an early domestic enterprise in the field of AI, Xinhua Group has long-term technology accumulation and practical experience. Facing the future, Xinhua San will adhere to the concept of "intensive cultivation and pragmatism, endow the times with wisdom", focus on AIGC technological innovation, actively explore industry application scenarios, build leading digital solutions with pragmatic ideas, accelerate the intelligent upgrading of digital change through the constantly evolving digital brain, and be the most reliable partner of Baixing and Baiye.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.