Module 12 Activity

Scenario

The deliverable the course has been building towards: a model, its evidence, and an honest card.

Task

Work through these in order. Do not start training before step 2 is written.

1. Frame the problem: decision, unit of prediction, target, what is known at prediction time, and

the cost of each error type.

1. Write the dataset card and split plan: identify the repeated unit and split on it.
2. Fit both baselines and the network, and report all three.
3. Log training evidence: per-epoch losses, gradient norms, best validation and its epoch.
4. Analyse errors and choose a threshold, recording why.
5. Write the model card, including out-of-scope uses and next steps.

Deliverable

Three files: the training script, the diagnostics record, and the model card.

Check your work

• A row-level split leaks learners when each appears several times. The grouped split has zero overlap.
• Report majority, logistic and network. If the network only beats the majority class, ship the simpler

model.

• epochs_trained_past_best should be small. If it is large, you discarded the best model.

Marking rubric

Mark complete