Phi Silica Task Specialization with LoRA: Technical Deep Dive

Key Points

• Microsoft announced support for LoRA finetuning on Phi Silica, a small language model running on Copilot+ PCs, enabling task-specific customization with minimal overhead.
• Using LoRA, Microsoft customized Phi Silica for generating Kahoot! quizzes in its new Learning Zone app, cutting rejection rates by 75% and improving quality scores by 4.6 times.
• The company used synthetic data, AI evaluation agents, and system prompt tuning to train and assess the model — results confirmed by both automated tools and human reviewers.

Microsoft has made advancements in on-device artificial intelligence for education, revealing how small language models can be tailored for specific tasks without sacrificing performance or speed.

At its Build 2025 conference, Microsoft announced support for LoRA (low-rank adaptation) finetuning on Phi Silica, a Small Language Model (SLM) that runs locally on Copilot+ PCs. LoRA enables lightweight customization by updating only a small part of the model, making it possible to perform specialized tasks efficiently using on-device AI.

This technology powers Microsoft Learning Zone, a new learning companion app for educators, developed under the code name “Project Spark.” Available at no cost, the app allows teachers to create interactive lessons and Kahoot! quizzes powered by AI — entirely on the device, without needing cloud processing.

To build this system, Microsoft collaborated with Kahoot!, a widely used educational platform, to generate classroom-ready content using Phi Silica + LoRA. Instead of training separate models for each task — which would require more memory and computing power — the team used LoRA adapters to retrain a single Phi Silica model for different quiz-generation needs.

Creating high-quality quizzes involved two main aspects: verifiable quality and subjective quality. Verifiable elements included following Kahoot’s format rules, such as maximum text lengths. Subjective qualities included how engaging, clear, and useful the questions were for students.

To train the model, Microsoft used GPT-4o to generate thousands of synthetic question-and-answer pairs based on educational content. These examples adhered to Kahoot!’s interface guidelines before being tested and refined using Phi Silica and LoRA. In total, 13,000 examples were created — 10,000 for training and 3,000 for testing.

Using Microsoft’s AI toolkit, LoRA adapters were trained to specialize Phi Silica for Kahoot-style quiz generation. Over time, these adapters allowed the team to shorten the system prompts required during task execution, making the process faster while still delivering accurate, structured outputs.

Improvements were also made to the model’s evaluation process. To measure quality at scale, Microsoft created an AI “review team” using a multi-agent system. Each AI agent represented a reviewer or critic, comparing the model’s outputs against defined quality criteria. This method helped identify strengths and weaknesses in the output, leading to better and more consistent results.

Human reviewers later validated the AI-generated assessments in a large-scale A/B test involving over 2,300 quiz pairs. Results showed that users preferred quizzes generated with Phi Silica + LoRA at a 4.6x higher rate, demonstrating clear improvements in clarity, correctness, and educational value.

Overall, the project reduced failed quiz creations by 75% while increasing perceived quality, proving LoRA to be a practical and scalable solution for personalizing small models in constrained environments like school classrooms.

Microsoft plans to release the Kahoot! quiz generator in public preview for educators later this summer, showcasing the potential of local AI in real-world applications.

The full technical report and open-source tools are available on Microsoft’s Build 2025 announcement page.

Read the rest: Source Link

You might also like: Try AutoCAD 2026 for Windows, best free FTP Clients on Windows & browse the best Surface Laptops to buy.
Remember to like our facebook and our twitter @WindowsMode for a chance to win a free Surface every month.