Building Agentic Solutions with Microsoft Foundry

Chapter 5. Model Customization Through Fine-Tuning

In this chapter, we will learn how to fine-tune models in Foundry using custom training datasets. Customizing a model with your own data can improve performance on specific tasks or domains by teaching the model to better understand your unique requirements and to respond with more accurate answers or in a desired style. Fine-tuning is a powerful technique that allows you to adapt pre-trained models to your specific use case without needing to train an entire model from scratch, which can be resource-intensive (expensive because you’d have to have many GPUs) and require large amounts of data. Instead, you can achieve significant improvements with just a few hundred or thousand examples that are relevant to your specific domain or task.

Fine-Tuning with Custom Data

When LLMs are trained, they learn from a broad range of data sources to develop general language understanding. Groups like Anthropic or OpenAI train their models on trillions of tokens across various domains from all sorts of sources including Common Crawl (an archive of web page data), Wikipedia, and digitized books. Thus, for giant models, they often respond reasonably well to a wide range of prompts without any additional training. You’ve probably also noticed that some alternate versions of models are released for specific use cases, such as code generation, deep reasoning, or mathematics. These models have been fine-tuned on additional data relevant to those tasks ...