Module 6 — Documentation: model cards and datasheets

Modules 2 to 5 produced numbers, plots and design decisions. This module gathers them into two documents — a model card and a datasheet for datasets — that make the audit legible to people who did not run it. Without those documents, the audit exists only in the head of the person who did it, and no future operator, regulator or user can act on it.

The two templates cover complementary ground. The datasheet describes the raw material (the training data). The model card describes the artefact built from it (the trained model). Neither replaces the other, and both need to live where the model is used, not in a wiki nobody reads.

Datasheets for datasets

Proposed by Gebru et al. (2018), a datasheet answers a fixed list of questions about the dataset's origin, composition, collection process, preprocessing, uses and maintenance. The point is to force the questions that always turn out to matter, in a shape that can be filled once and reused.

The seven sections that matter most:

Motivation — Why was the dataset created? Who paid for it? What tasks was it explicitly designed to support, and which ones was it not?

Composition — What does one row represent? How many rows and columns? Are there missing values? Are there subgroups (protected classes) and what are their proportions?

Collection — How was the data gathered? Over which time window? Were the individuals aware their data was collected? Was consent recorded? Which populations are excluded?

Preprocessing — What cleaning, filtering, imputation, feature engineering was applied? Is the raw dataset preserved? Are the preprocessing scripts versioned?

Uses — What tasks has the dataset been used for so far? Are there tasks for which it should not be used, and why?

Distribution — Who has access? Under what license? Are there export controls?

Maintenance — Who is responsible for updates? How are errors reported and corrected? When will the dataset be deprecated?

For the credit dataset, the "Composition" section states that one row is a loan application filed between 2013 and 2024, that gender is present, that occupation and ZIP code act as proxies for gender and ethnicity (finding of module 2), and that no applications from customers under 21 or over 75 are included. The "Uses" section explicitly bans training a marketing model on this dataset: the informed consent covered credit decisions only.

Model cards

Model cards, proposed by Mitchell et al. (2019), do the same job for the trained model. A model card fits on a couple of pages and answers:

Model details — Owner, version, date, training framework, license.

Intended use — What the model is for. What input distribution it expects.

Out-of-scope uses — What it is not for. This section prevents most misuse: if the card explicitly says "not to be used to price insurance" and someone uses it to price insurance, the accountability is clear.

Metrics — Which metrics were chosen and why. Numbers on the held-out set, broken down by subgroup. For the credit scorer: accuracy, AUC, selection rate, equalized odds difference, calibration by decile, per group.

Training data — Pointer to the datasheet.

Evaluation data — Sometimes different from the training set; e.g. a validation set from a different time window to test drift.

Ethical considerations — Known risks, mitigations, residual disparities.

Caveats and recommendations — Where the model has not been tested, what monitoring is required in production.

A concrete extract from the credit scorer's model card

Intended use. Score applications for standard consumer loans between EUR 3 000 and EUR 50 000, twenty-four to sixty months, filed at branches or via the online portal, from applicants aged 21 to 75, residing in the seven regions where the training data was collected.

Out of scope. Mortgage decisions, business loans, credit card issuance, applications from customers outside the seven training regions, pricing (the model outputs a probability, not an interest rate — pricing is a separate policy). Any use in an automated-only mode without the human reviewer of module 7.

Metrics (held-out 2024 Q3, N = 12 800). Overall AUC 0.87. Selection rate 0.71 for men, 0.63 for women, 0.68 for 21–35, 0.73 for 36–55, 0.62 for 56–75. Equalized odds difference by gender: 0.028 (below the committee threshold of 0.03). Equalized odds difference by age band: 0.047 (below 0.05). Calibration expected calibration error: 0.019.

Ethical considerations. The model uses ZIP code as a feature; module 4 shows it accounts for 0.14 of a typical rejection, and module 2 shows it is a proxy for gender and ethnicity. The committee chose to keep it because removing it drops AUC by 0.06 with only a 0.005 improvement in the fairness metric. This trade is revisited quarterly.

Caveats. The model has not been tested on applicants who were previously rejected and are re-applying. Monitoring: monthly recomputation of the two fairness metrics, quarterly recomputation of calibration, drift alarm when input feature distributions shift by more than 5 % KS distance.

Why the "intended use" line matters most

Every scandal that reaches the press has the same structure: a model was trained for task A and used for task B without the users knowing. The intended-use and out-of-scope-uses sections of the model card are the single most effective defense, because they turn scope creep from an ambiguity into an explicit violation. Any downstream integration that operates outside the declared use must obtain sign-off from the risk committee, on record.

The model card is not a marketing document. It contains numbers that competitors will read. It contains admissions of residual bias that headline writers may exploit. Publishing it anyway is what a responsible organization does; the alternative — an undocumented model — leaks worse and more damaging information the first time something goes wrong.

Summary

• The datasheet describes the dataset, the model card describes the trained model — both are needed, and both are useless in a wiki nobody reads.
• The intended use and out-of-scope uses sections prevent scope creep and shift accountability when misuse happens.
• Metrics are reported by subgroup with the disparity numbers made explicit, not buried under an aggregate score.
• Ethical considerations must record known residual bias and the deliberate trade-offs behind it, not paper over them.

Next module: the human in the loop, the biases of automation, and the recourse procedure a rejected applicant can invoke.