Fine-tune a language model

Why it matters

Fine-tuning is the costliest fix for a quality gap, so it has to beat a well-written prompt on the same task before it earns its place.

What to practise

Start a supervised fine-tuning job in Microsoft Foundry on a JSONL file of chat examples, with the result deployed automatically to a Developer deployment. While it runs, refine a base model’s instructions in the playground. Then give the fine-tuned model the same instructions and questions, and compare its consistency and conversational tone with the base model’s. Read the training file: every example repeats the system message and shows the target response style. The exercise notes that some of its features are in preview.

How it fits

Use it for 2.1.6: the gap here is style consistency, and supervised fine-tuning is the method chosen for it. The comparison is informal, though; 2.1.7 asks for held-out evaluation and governed data before a fine-tuned model replaces the prompt.