A general-purpose model may understand a customer review yet miss the vocabulary of a specialist bookshop. Fine-tuning trains a chosen base model on labeled examples so its recommendations better reflect that vocabulary. Try it when an off-the-shelf model falls short—not simply because a bigger model is available. Compare quality and fairness alongside cost and latency before deciding whether to fine-tune or switch models. section 7.4.3
In Azure AI Foundry, the workflow is:
- Prepare a JSON Lines (JSONL) file with one input–output pair per line, such as a review and the recommendation it should produce. section 7.4.4
- Open Fine-Tuning in an accessible AI Foundry workspace, start the wizard, and choose a base GPT model. Select the JSONL file from linked Blob Storage or upload it from your workstation. section 7.4.4
- Review the settings and submit the job. Inspect its status and recorded training details; when satisfied with the result, deploy the model as an endpoint for application requests. section 7.4.4
Listing 7.1 Example JSONL file for Banning Books
{"input": "I loved the intricate world-building and moral dilemmas in 'The
Handmaid's Tale'. Margaret Atwood's prose keeps me hooked.", "output":
"Recommend 'Never Let Me Go' by Kazuo Ishiguro—another speculative novel
with deep ethical questions and beautiful writing."}
{"input": "The coding examples in 'Learning Python' were clear, but I wish
there were more real-world projects.", "output": "Suggest 'Fluent Python'
by Luciano Ramalho—it offers advanced, hands-on Python projects to level up
your skills."}
{"input": "This hacker thriller had me on the edge of my seat, but the tech
jargon was a bit dense for a newcomer.", "output": "Recommend 'Zero Days'
by Kim Zetter—a gripping cybersecurity exposé that's more accessible for
readers new to hacking topics."}
{"input": "I appreciated the historical context in 'The Code Book', but the
math sections were challenging.", "output": "Suggest 'The Golden Ticket' by
Lance Fortnow—a computer science history that focuses on stories rather
than heavy mathematics."}
{"input": "The practical examples in 'Azure in Action' helped me deploy
services quickly, but I'd like a deeper dive into AI features.", "output":
"Recommend 'Architecting Microsoft Azure AI Solutions' by Drobny and
Perfekt—it offers in-depth coverage of Azure AI services with hands-on
labs."}
The sample JSONL pairs customer-review inputs with desired book recommendations, making the training format concrete. section 7.4.4
A deployed endpoint is not the end of the decision. Keep weighing reader satisfaction and operating cost, and monitor real-world performance and fairness as the recommendation feature moves toward production. section 7.4.3 section 7.6
