Twenty-two questions, two from each of the eleven lessons. They run in order: Foundations (1โ4), Self-Hosted Track (5โ6), Managed Track (7โ8), then Quality and Delivery (9โ10), and finally Strategy (11). Answer all of them to see your grade.
๐ง Grand Quiz
1. Your model keeps inventing product prices that changed last week. What is the correct fix?
2. Which of these is the strongest genuine justification for fine-tuning?
3. "We trained with QLoRA and then ran DPO." Which two independent axes do those terms describe?
4. What is the practical difference between shipping a LoRA adapter and shipping a merged model?
5. Why is a full fine-tune of a 7B model so much more memory-hungry than QLoRA, even before activations?
6. In W = Wโ + (ฮฑ/r)ยทBยทA, you raise rank from 16 to 64 but leave alpha at 16. What happens to the effective update scale?
7. You have 300 real inputs and no labels. What is the standard highest-ROI way to build a training set?
8. Your fine-tune shows excellent training loss, but the deployed model ignores instructions and produces odd turn structure. Most likely cause?
9. What does load_in_4bit = True buy you in Unsloth, and at what cost?
10. Unsloth's docs tell you to adapt all seven linear modules rather than attention only. Why?
11. You have a free 16 GB T4 and want to fine-tune an 8B model. Which method, and why?
12. With per_device_train_batch_size=2 and gradient_accumulation_steps=8, what is the effective batch size, and which single change saves VRAM?
13. Which file requirement is unique to the Azure AI Foundry path and commonly causes validation failures?
14. You specifically want to fine-tune o4-mini on Azure. Which method must you use?
15. A fine-tuned Azure chatbot costs $1,422/month for 10,000 conversations. What is the largest component?
16. Which single change cuts monthly hosting cost the most on Azure?
17. What is the one thing you must record before you start training?
18. Training loss keeps dropping while validation loss rises. What is happening?
19. What does merging a LoRA adapter actually do?
20. You must serve five fine-tunes of the same 8B base model. Cheapest approach?
21. What is the structural cost difference between self-hosting and a managed endpoint?
22. What should you do first when you have 500 examples and don't yet know if the task is learnable?