Module 11 Activity

Scenario

Everything the model learned, it learned from decisions people made while collecting.

What you build

A bias and privacy assessment with per-group results and a model card whose prohibitions are specific.

Steps

1. List six collection decisions and what each embedded. Include what was discarded and why.
2. Define groups from collection metadata - shift, device, site, lighting - before measuring anything.
3. Report accuracy and the metric that matters per group, with row counts.
4. Sort five proposed uses into intended, unvalidated and prohibited, with a reason each.
5. Write the model card with the baseline, per-group results and known gaps including any label-agreement figure.
6. Use three permission categories, and give every 'not yet' a stated condition.

Evidence to hand in

• The six collection decisions and what each embedded.
• Groups defined before measurement.
• Per-group results with row counts.
• The five uses sorted with reasons.
• The model card.
• The three-category permission list with conditions.

Review checklist

• The discarded-images decision is examined, not just the kept ones.
• Groups were defined before results were seen.
• Any label-agreement figure appears in the card.
• 'Not yet' entries state what would have to be true.
• Capability is not treated as permission.

Mark complete