Fine-tune Phi-3.5 models in Azure by info.odysseyx@gmail.com November 1, 2024 written by info.odysseyx@gmail.com November 1, 2024 0 comment 1 views 1 that Pi Model Collection It represents the latest advancement in the Microsoft Small Language Models (SLM) series. In August 2024 we welcomed our latest addition. pi-3.5-mini and Phi-3.5-MoEMix of Experts (MoE) Model: Phi-3.5-mini: This 3.8B parameter model provides multilingual support, enhanced inference capabilities, and an extended context length of 128K tokens. Phi-3.5-MoE: With 16 experts and 6.6 billion active parameters, the model offers high performance, reduced latency, multilingual support, and strong safety measures, outperforming the capabilities of larger models while maintaining the efficiency of the Phi model. From general users to custom SLM Benchmark results highlight its incredible efficiency and capabilities. pi-3.5-mini and P-3.5-MOE. However, you can nonetheless further customize the model to suit your unique requirements in order to match the performance of larger models for specific tasks. There are three powerful techniques you can use to customize your language model to fit your organization’s specific needs. Rapid Engineering Search Augmentation Generation (RAG) fine tuning Let’s take a closer look at each of these technologies. Rapid Engineering The idea is to provide clear instructions directly within the prompt, often from a system prompt, to guide the model’s response. This method falls into the category of “providing additional information to the prompt” and can be particularly useful for shaping the behavior and output format of your model. Next we Create search augmentationor piece. This technique is used when you want to integrate your organization’s data and knowledge into the model’s responses. RAG allows you to provide your model with a trusted source of answers through additional documentation. This improves the model’s ability to retrieve relevant information and reinforce prompts to generate informed, contextually accurate responses. RAG also falls into the “Provide additional information when prompted” category. fine tuning The process of customizing a model using labeled training data, often resulting in improved performance and reduced computational resources. In fact, fine-tuning smaller models with appropriate training data can outperform larger models for certain tasks. In particular, Low-Rank Adaptation (LoRA) fine-tuning is an excellent approach to tailor language models to specific use cases due to several key advantages: First, LoRA significantly reduces the number of trainable parameters, making the fine-tuning process more efficient and saving time and money. This reduced demand on resources speeds up iteration and makes it easier to experiment with fine-tuning operations in LoRA. LoRA keeps the original model weights almost unchanged, helping it adapt to specific tasks while maintaining the general functionality of the pretrained model. Fine-tuning LoRA for Phi-3.5 models Today we are proud to announce that LoRA fine-tuning is now available. pi-3.5-mini and Phi-3.5-MoE Models will be available in Azure AI starting November 1, 2024. Serverless fine-tuning for Phi-3.5-mini and Phi-3.5-MOE models allows developers to quickly and easily customize models for cloud scenarios without having to manage compute. The fine-tuning environment is available in Azure AI Studio and takes a pay-as-you-go approach, so you pay only for the actual training hours required for fine tuning. Once fine-tuned, you can deploy your model to Azure for inference with the option to enable content safety. Deploying fine-tuned models is a streamlined process with pay-as-you-go services. Fees for deploying a fine-tuned model are billed based on the number of input and output tokens used, along with a nominal hosting fee for maintaining the fine-tuned model. Once deployed, you can easily enhance your AI capabilities by integrating your fine-tuned models with leading LLM tools such as Prompt Flow, LangChain, and Semantic Kernel. Fine tuning is also possible through the Managed Compute option. You can use the Azure Machine Learning Studio user interface or follow these: laptop example Create custom models. The advantage of using a laptop is that it allows more flexibility in the configuration used for fine-tuning operations. You have the option to download a fine-tuned model and then deploy it using Azure managed compute resources, your own premises, or edge devices. The fine-tuned model is also licensed under the MIT License. closing remarks The Phi-3.5 model family represents a significant advance in the area of SLM. We push the boundaries of performance, efficiency, and customization with the introduction of Phi-3.5-mini and Phi-3.5-MoE. The availability of LoRA fine-tuning in Azure AI allows developers to tailor these models to their specific requirements to ensure optimal performance for a variety of applications. We are committed to continuously innovating and improving our models while providing cutting-edge solutions that drive progress and enhance the user experience. Thank you for joining us on this journey. We look forward to the amazing results we will achieve with the Phi-3.5 model. Source link Share 0 FacebookTwitterPinterestEmail info.odysseyx@gmail.com previous post A year of innovation- and feedback-driven features in Microsoft Word next post Cumulative Update #29 for SQL Server 2019 RTM You may also like AI search threatens digital economy, researcher warns November 12, 2024 Qualcomm has an ‘AI-first’ vision for the future of smart devices November 11, 2024 AMD is moving fast in AI, may join forces with Intel November 11, 2024 A New Dawn of Software Defined Networking (SDN) in Windows Server 2025 November 5, 2024 Get AI ready: Empowering developers in the era of AI November 5, 2024 Announcing the General Availability of Windows Server IoT 2025! November 5, 2024 Leave a Comment Cancel Reply Save my name, email, and website in this browser for the next time I comment.