Exploring Small Language Models: From Theory to Practice
Join Alon Fliess, CTO of Zionet, Microsoft AI Platform MVP, and Regional Director, for a comprehensive session exploring the transformative potential of Small Language Models (SLMs). As the demand for cost-effective, flexible, and efficient AI solutions grows, SLMs are emerging as a powerful alternative to large language models. This lecture will explore the capabilities of SLMs, focusing on the Phi model family and its practical applications.
The session will begin by examining the evolution of language models and the growing need for smaller, more efficient models. We will then discuss the advantages of SLMs, including improved cost efficiency, enhanced security and privacy, deployment flexibility, and reduced latency. Attendees will gain insights into the Phi model family, highlighting its technical specifications, strengths, and best-fit scenarios.
We will then demonstrate how to run Phi models locally, showcasing their performance in offline environments, on-device solutions, and resource-constrained systems. This practical demonstration will emphasize the advantages of local inference, particularly in latency-critical or cost-constrained scenarios. Additionally, we will showcase Llama2.c running on an ESP32, demonstrating how a tiny language model can generate short stories on an extremely resource-restricted device, showcasing the potential for AI in embedded systems.
In the final segment, we will introduce Azure AI Foundry and demonstrate how to fine-tune Phi models for specialized use cases. Attendees will learn best practices for tailoring models to meet domain-specific requirements, ensuring improved accuracy and performance in targeted applications.
By the end of this session, participants will have a comprehensive understanding of Small Language Models, practical insights into running and optimizing Phi models, and the skills to fine-tune them using Azure AI Foundry. The lecture will combine theory with practical demonstrations, ensuring attendees leave with actionable knowledge to apply to their AI development projects.