Post-Training and Deploying Open Source Reasoning Models in Foundry | BRK232

From the creator

Open-source reasoning models are powerful out of the box, but production performance comes from closing the loop. Learn how to use Microsoft Foundry to collect production traces, curate them into datasets, and post-train reasoning models with RL using frameworks like slime, verl, and TRL. We'll cover when RL drives real gains and how to redeploy improved models without managing infrastructure. Seating for this session is first-come, first-served. Add it to your schedule to plan your day and arrive early to secure a spot. To learn more, please check out these resources: * https://aka.ms/build26/BRK232 * https://aka.ms/build/foundrydiscord 𝗦𝗽𝗲𝗮𝗸𝗲𝗿𝘀: * Chris Lauren * Manoj Bableshwar * Vijay Aski 𝗦𝗲𝘀𝘀𝗶𝗼𝗻 𝗜𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻: This is one of many sessions from the Microsoft Build 2026 event. View even more sessions on-demand and learn about Microsoft Build at https://build.microsoft.com BRK232 | English (US) | Working with models Breakout | (300) Advanced #MSBuild Chapters: 0:00 - Advantages of fine-tuning for reducing costly tool interactions 00:05:22 - Demonstration of Foundry using a retail customer service agent scenario 00:14:35 - Post-Training Objectives: Task Optimization and Agent Workflow 00:15:00 - Sources of Data for Post-Training: Traces, Human Preferences (RLHF), and Synthetic Data 00:19:00 - Reinforcement Fine-Tuning Explained: Reward-Based Learning for Verifiable Tasks 00:24:20 - Setup completion and starting SFT job 00:35:05 - Monitoring Training Checkpoin

Published 2026-06-04 · Watch on YouTube ↗
Availability and playback restrictions are controlled by the original channel.