微软公司于7月25日发布博文，宣布在Azure云服务上支持开发者微调（fine-tune）Phi-3-mini和Phi-3-medium两个AI模型，此举旨在提升模型在特定应用场景中的性能。开发者可以根据不同需求，通过微调Phi-3-medium模型，为学生提供个性化的辅导，或构建具有特定语气或响应风格的聊天应用。Phi-3-mini模型于今年4月发布，拥有38亿个参数，提供4K和128K两种上下文长度版本；而Phi-3-medium模型参数量高达140亿，同样支持4K和128K上下文长度。
在性能方面，Phi-3-mini模型在6月份的更新后，性能显著提升，具体指标如下：在“Instruction Extra Hard”测试中，从5.7提升至6.0；“Instruction Hard”测试中，从4.9提升至5.1；“JSON Structure Output”测试中，从11.5跃升至52.3；“XML Structure Output”测试中，从14.4增长至49.8；“GPQA”测试中，从23.7进步到30.6；“MMLU”测试中，从68.8增加至70.9。平均性能指标从21.7提升至35.8。
微软公司今日还宣布推出Models-as-a-Service（serverless endpoint）服务，使开发者能够轻松构建基于Phi-3-small模型的人工智能应用，无需关注底层基础设施。未来，微软计划进一步扩展服务，推出包含图像处理能力的Phi-3-vision模型。这一系列举措旨在简化AI模型的应用开发流程，降低技术门槛，加速AI技术的普及与应用。
此外，IT之家提醒用户，文中包含的对外链接旨在提供更多信息，用户在点击链接时应谨慎判断，确保链接内容的可靠性。IT之家始终遵循透明原则，确保所有文章内容的准确性和公正性。
通过微软在AI模型微调和云服务上的创新举措，AI技术的应用边界有望进一步拓展，为开发者和企业提供更多可能性，促进人工智能技术在教育、交互式应用等领域的创新应用。
英语如下：
Headline: “Microsoft Opens Up Fine-Tuning of AI Models Phi-3-mini, Phi-3-medium to Boost Personalized Development”
Keywords: Microsoft support, AI models, developer fine-tuning
Content: On July 25, Microsoft announced in a blog post that it is now supporting developers to fine-tune (also known as ‘tuning’) the Phi-3-mini and Phi-3-medium AI models on its Azure cloud service. This initiative aims to enhance the models’ performance in specific application scenarios. Developers can customize the Phi-3-medium model to offer personalized tutoring to students or build chat applications with tailored tone and response styles. The Phi-3-mini model, released in April with 3.8 billion parameters, offers versions with context lengths of 4K and 128K. The Phi-3-medium model, with a staggering 14 billion parameters, also supports the same context lengths.
In terms of performance, following its update in June, the Phi-3-mini model showed significant improvements. Here are the specific metrics:
– In the “Instruction Extra Hard” test, the score improved from 5.7 to 6.0.
– In the “Instruction Hard” test, the score went from 4.9 to 5.1.
– The “JSON Structure Output” test saw a leap from 11.5 to 52.3.
– In the “XML Structure Output” test, the score rose from 14.4 to 49.8.
– The “GPQA” test saw a progress from 23.7 to 30.6.
– In the “MMLU” test, the score increased from 68.8 to 70.9.
The average performance indicator soared from 21.7 to 35.8.
Additionally, Microsoft announced the launch of Models-as-a-Service (serverless endpoint) service, enabling developers to easily build AI applications based on the Phi-3-small model without having to manage the underlying infrastructure. Looking ahead, the company plans to further expand its services, including the Phi-3-vision model for image processing capabilities. These moves are geared towards simplifying the development process for AI models, reducing technical barriers, and accelerating the widespread adoption and application of AI technology.
Furthermore, IT House reminds users that the external links included in the text are provided for additional information. When clicking on these links, users should exercise caution to ensure the reliability of the content. IT House adheres to transparency principles, ensuring the accuracy and fairness of all article content.
Through its innovative approach to AI model fine-tuning and cloud services, Microsoft is poised to extend the boundaries of AI technology application, offering developers and enterprises more possibilities, and driving innovation in AI applications across sectors like education and interactive applications.
【来源】https://www.ithome.com/0/784/434.htm
Views: 2