A model card is the README for a model: what it does, what it was trained and evaluated on, where it works, and — most importantly — where it doesn't. It turns an opaque artifact into something a reviewer, a downstream team, or a regulator can actually assess. Write it before launch, not after.
Required Inputs
Ask for these only if they aren't already provided:
- Model name & version, owner team, and date.
- What it does — task type (classification, generation, ranking, extraction…) and the decision it informs.
- Intended use & users — the supported use cases, and explicitly the out-of-scope ones.
- Training data — sources, size, time range, and known gaps (link a
dataset-datasheetif one exists). - Evaluation — datasets, metrics, and results, ideally broken down by subgroup/slice.
- Known limitations & risks — failure modes, bias findings, safety concerns.
Output Format
Model Card: [name] v[version]
Owner: [team] · Date: [date] · Status: [in review / production / deprecated]
1. Overview — one paragraph: what the model does, the decision it serves, and who uses it.
2. Intended Use
- In scope: the use cases this model is validated for.
- Out of scope / do not use for: explicit prohibited or unvalidated uses (this section prevents the most harm).
- Users: who is expected to operate or consume it.
3. Training Data — sources, size, time window, labelling method, and known coverage gaps.
4. Evaluation
- Metrics: the primary metric(s) and why they were chosen for this task.
- Overall results: headline numbers vs. a stated baseline.
- Sliced results: a table of the key metric across important subgroups (geography, language, device, demographic where appropriate) — surface where performance drops, don't hide it behind an average.
| Slice | N | Metric | vs. overall |
|---|
5. Limitations & Failure Modes — concrete situations where it underperforms or should not be trusted.
6. Ethical Considerations & Bias — fairness findings, sensitive-attribute handling, and mitigations applied.
7. Deployment & Monitoring — serving constraints (latency/cost), the drift/quality signals you'll watch, and the rollback trigger.
Quality Checks
- "Out of scope / do not use for" is filled in with specifics — not left blank
- Evaluation is reported by slice, not just one global average that hides subgroup harm
- Every metric states the baseline it's measured against
- Limitations describe real, concrete failure situations (not "the model may be imperfect")
- A monitoring signal and an explicit rollback trigger are named
Anti-Patterns
- Do not report a single aggregate metric and call evaluation done — averages mask the slices where a model fails worst
- Do not leave "intended use" open-ended — an undefined boundary is an invitation to misuse
- Do not omit known biases because they're uncomfortable — an undocumented risk is a worse liability than a documented one
- Do not present accuracy without the class balance / base rate — 95% accuracy on a 95/5 split is meaningless
- Do not ship without a monitoring plan — a model card without a rollback trigger is a snapshot, not a contract
Based On
Model Cards for Model Reporting (Mitchell et al., 2019) and the model-documentation practice used in responsible-AI reviews.