Weekly Machine Learning Roundup: Fabric ML and LLM Fine-Tuning
This week’s ML coverage spotlights Microsoft Fabric’s expansion of analytics and machine learning, with practical routes for ML in production, fine-tuning workflows, and automated pipelines. Tools like Semantic Link and Foundry fine-tuning offer easier, AI-driven analytics and operational intelligence.
Microsoft Fabric for ML, AI, and Operational Analytics
Microsoft Fabric now better unifies analytics, machine learning, and business reporting. Semantic Link is now generally available, allowing a shared semantic layer for data engineering, AI, and BI to use common models. It supports semantic model updates directly from notebooks, immediate sync to Power BI, and harmonized workflows. Automation is easier with tighter SQL/Spark orchestration, while community repositories provide reusable patterns. For IoT and streaming data, Fabric’s operational analytics uses time series dashboards with Kusto, dynamic slicing, anomaly detection, and DirectQuery for live reporting. These updates expand the platform’s ability to handle large-scale, real-time data. ML workflows in Fabric and Power BI are progressing, letting teams run predictions in dashboards using LightGBM/SMOTE, OneLake-backed data, and MLflow for automation. The Fabric IQ platform provides the foundation for digital twins and ontologies, supporting smarter knowledge and automation development.
- Supercharge AI, BI, and Data Engineering with Semantic Link in Microsoft Fabric
- Adaptive Time Series Visualization at Scale with Microsoft Fabric
- Integrating Machine Learning with Power BI Reports in Microsoft Fabric
- Fabric IQ Overview
Fine-tuning and Preference Optimization for Large Language Models on Azure
A hands-on guide is available for fine-tuning enterprise LLMs with Microsoft Foundry on Azure, taking models and aligning them for organization-specific requirements and policies. The documentation covers data prep, running training jobs, and benchmarking—applying methods to use cases like PubMed summarization for health and science. Direct Preference Optimization (DPO) is also explained, detailing how human feedback can steer LLMs toward better outputs. DPO is now in the Foundry SDK, and tutorials include example code, best practices for parameter selection, and links to new documentation.