Module 12 Activity

Scenario

The capstone: a responsible vision workflow, scoped, measured per slice, and documented with its limits.

What you build

A complete vision project - scope, dataset record, baselines, per-slice evaluation and a model card with more prohibitions than permissions.

Steps

1. Scope by the decision the model supports, state the error-cost asymmetry as a ratio, and make refusal conditions computable from the image.
2. Record the dataset: schema rulings, class counts, grouped split, near-duplicate check and label audit.
3. Compute the majority baseline and a trivial baseline before training.
4. Report the model against both, on the metric that matters rather than on accuracy.
5. Evaluate per slice with row counts, and mark small slices unmeasured.
6. Write the model card with not_approved_for, accepted limits with residual risk, untested areas and a review trigger.

Evidence to hand in

• The scope with its four checks.
• The dataset record including the label audit.
• Both baselines computed before training.
• The three-way comparison on the metric that matters.
• Per-slice results with usability verdicts.
• The model card.

Review checklist

• Refusal conditions are computable from the image, not model-dependent.
• The label audit happened before training.
• Classes with too few examples are reported, not measured.
• The deployment decision is per slice, not overall.
• not_approved_for is longer than intended_use.
• There is a stated review trigger.

Mark complete