Validate and launch your customized models seamlessly using Microsoft Foundry! In this installment, we tackle what happens after the training is complete. You'll learn how to deploy fine-tuned models effectively, manage long-term inference costs, and rigorously evaluate your model's performance post-training using a custom grader logic to ensure production-grade reliability.

00:03 Welcome and scenario
00:52 Post-training evaluation
01:20 Demo: Using a custom grader for evaluations
03:35 Cost management
05:50 Model deployment

Microsoft Foundry - https://aka.ms/foundry-ft
Foundry Finetuning Demos on GitHub - https://aka.ms/ft-demos
Azure OpenAI Fintuning Costs: https://aka.ms/aoai-ft-cost

Bethany Jepchumba, Twitter/X - https://twitter.com/bethanyjep
Bethany Jepchumba, LinkedIn - https://www.linkedin.com/in/bethany-jep/
Bethany Jepchumba, GitHub - https://github.com/bethanyjep

Microsoft Developer

A walkthrough of model optimization in Microsoft Azure AI Foundry covering three key areas: evaluation, deployment, and cost management. Evaluation uses a custom Python grader with precision, recall, and F1 score to compare GPT-4.1, GPT-4.1 mini, and a fine-tuned model on agentic tool calling tasks, showing a ~6% improvement after fine-tuning. Deployment options include global, regional, and developer tier (free but limited to 24 hours). Cost considerations cover training tokens, hosting charges, and regional pricing differences for both supervised fine-tuning and reinforcement fine-tuning models.

Model Optimization in Microsoft Foundry: Deployment and Evaluations