Class 10 · Year project
Build it, measure it, answer for it
All year you've learned to build models, measure them and question them. Now build one that matters — and stand up in front of the people it's for.
The big idea
In Class 9 you proposed a model. This year you build one, measure it honestly, and answer for it in front of the people it's meant to serve.
Review it → build it → test it on data it never saw → find who it fails → fix what you can → say plainly what you couldn't.
Two things make this project serious rather than a poster:
- The test data is real and unseen. Kept aside from the start, collected where the model will really be used.
- You say who it fails. Every model fails someone. A project that finds them and says so is worth more than one that claims it works for everyone.
📱 Scan to watch: From a review to a model you can defend —
Your group's project
Start from your ethics review from Lesson 1.5 (revised if it said change it). Your model is a classifier — it sorts something into classes — built with one of:
- Teachable Machine — images or sounds (Units 2 and 4)
- Orange — a table of data (Unit 2)
- Python — text, with a word list or keywords (Unit 5)
It must serve a Sustainable Development Goal. Name it on page one.
Before period 1 (two weeks of homework)
- Collect your data. Training data and, kept completely separate, testing data — collected somewhere else, or at a different time, the way the model would really meet it. At least 20 test examples.
- Ask two real users what they would need the model to do, and what a mistake would cost them.
- No personal data about anyone. If your data is about people, your ethics review must show the consent and why you need it — or your project should change.
Period 1: Build it
- Revise your ethics review (page 1 of the report).
- Build the model on the training data only.
- Write "How it decides" in plain words (Unit 2 project): what it looks at, and how.
Period 2: Measure it
- Test on the testing data only. Build the confusion matrix, with the positive class named.
- Calculate accuracy, precision, recall, F1. Choose the metric that matters — false alarm or miss, which does more harm to your users?
- Split it by group: place, season, light, device, language — whatever could matter here. Find who it fails.
Period 3: Fix what you can, and write it down
- Make one improvement aimed at the group it fails. Retest on the same test data.
- Write the model card (Unit 3): what it does and must not be used for · data · how tested · results overall and by group · weaknesses · who is accountable.
- Finish the report: ethics review · how it decides · evaluation · the improvement · model card · what we couldn't fix, and what we'd need.
Period 4: Answer for it
- Each group presents for 5 minutes, to the class and, where possible, to parents and the people it would serve: the problem · a live demo · the numbers · who it fails · what you'd do next.
- Then 3 minutes of questions. Every member answers at least one.
How it is marked
Marked on this rubric, which you see before you begin. Rows 1–7 are the group's; rows 8 and 9 are each person's own. Marks are private and never compared with anyone else's.
Total: 27. No row rewards a high accuracy. Every row rewards honesty.
Your log (row 9): a few lines after each period in your own folder — what you did, what went wrong, what you learned. It shows your part in a group project.
1. Ethics review
- 0Missing
- 1Copied from Lesson 1.5 unchanged
- 2Revised, with a real proxy and consent check
- 3Revised, and it changed the project
2. The model works
- 0No model
- 1Works on training data
- 2Works on unseen test data
- 3Works on test data collected where it would be used
3. How it decides
- 0Missing
- 1"It uses AI"
- 2Stated
- 3Plain sentences a user could check
4. Evaluation
- 0Missing
- 1Accuracy only
- 2Matrix and all four metrics
- 3Matrix, metrics, and the metric that matters, argued
5. Who it fails
- 0Not checked
- 1Split, not compared
- 2Split and compared
- 3Split, compared, explained, and the improvement retested
6. Model card
- 0Missing
- 1Partly done
- 2Complete
- 3Complete, honest, and useful to a real user
7. The presentation
- 0Not given
- 1Read from the page
- 2Clear, with a demo
- 3Clear, a demo, and honest about who it fails
8. Your question
Just you- 0Didn't answer
- 1Answered
- 2Answered clearly
- 3Answered clearly, with evidence from the project
9. Your log
Just you- 0Missing
- 1A few lines
- 2What you did each period
- 3What you did, what went wrong, what you learned