Build examples that teach the intended behavior
Examples teach both your intent and your mistakes
Suppose half your examples interpret “under $25” as item price and half as all-in cost. Training on both without a rule does not resolve the ambiguity; it teaches conflicting conventions. Define the desired behavior before collecting more data.
Each example needs an input, an expected output, and enough context to understand why the output is correct. Include corrections, unknowns, and edge cases rather than only perfect requests.
Check your understanding
Two almost identical versions of one conversation are in training and test. What is the concern?
Keep evaluation examples outside training
A training set teaches the behavior. A validation set helps choose settings. A final test set estimates performance on held-out examples. Near-duplicate conversations across these sets can create misleading confidence.
Split related examples together, such as variations of the same conversation. Remove unnecessary personal information and use authorized data. For this course, create fictional requests and keep your own final test cases out of prompt examples as well.
Check your understanding
Match each dataset to its role.
Review the label before blaming the model
Take “No meat, but fish is okay.” Labeling it vegetarian loses the exception. A reviewer should preserve what was actually said, or ask for clarification when the format cannot represent it.
Write three examples in your reflection, then have a second pass challenge them. Are the expected fields supported? Do they encode the budget basis? Do they distinguish unknown from none? Label quality is a product decision before it is a training input.
Check your understanding
Write a useful edge-case training example for the dinner assistant.
Build checkpoint
Open the Adaptation Lab and follow the practice above. Each lab explains its inputs, its actual logic, and its limits. Save your observations below; your check-ins and reflection remain in Product Academy.
For an AI-assisted code change, download the lab bundle from the lab page. Ask your coding assistant to explain the relevant function first, change one behavior, and add one test that could fail. Compare the result with your prediction.
Save your experiment and check-in
Record your prediction, what you tried, what happened, and the evidence for your conclusion. Name one thing you can now do independently and one uncertainty to revisit.
Key takeaways
- Treat labels as a behavioral specification.
- Separate training and evaluation cases.
- Include ambiguity and corrections, not just easy requests.