Preparing and Uploading Data

Get your prediction data into the right shape for a bias report.

The bias report analyses your model's real predictions. Garbage in, garbage out — here's the shape it needs.

Format

A CSV file with a header row. One row per prediction. You'll need at least:

  • a prediction/outcome column — what the model decided (e.g. hired, approved, score),
  • one or more protected attribute columns — the groups you want to compare (e.g. gender, age_band, nationality),
  • ideally a ground-truth column — what the correct outcome actually was. Without it, metrics that need true outcomes (like equalized odds) can't be computed, and you'll get parity metrics only.

Uploading

  1. Open Dashboard → Upload / Bias Report.
  2. Drop in the CSV — you'll see a preview of the detected columns and row count.
  3. Select which column is the prediction and which are protected attributes.
  4. Generate the report.

Data hygiene and privacy

  • Upload the minimum needed: predictions, attributes, outcomes. Leave names, emails, and IDs out — pseudonymise before export.
  • A few hundred rows per group is a practical minimum; metrics on tiny groups are noise.
  • Sample from production, not from a demo dataset — the report is only evidence if it reflects reality.

Bias reports are available on Starter (5/month) and unlimited on Pro. Your trial includes the Pro allowance.

Didn't find what you need?

Guardia AI provides compliance tooling, not legal advice. For official regulatory text, see EU Regulation 2024/1689.