Model Card
A model card documents a specific model's intended uses, evaluation evidence, limitations, and operating conditions so others can assess whether to use it.
Download Markdown (.md)Soroush’s Product Wiki · https://wiki.soroush.uk/topic/model-card/
01What it means
A model card is a structured report accompanying a trained model and its version. It connects intended and unsuitable uses with training information, evaluation methods, performance across relevant groups and conditions, and known limitations. It describes the model as a reusable component; a system card may cover the wider application, including tools and safeguards. Neither document is a certification.[1][2]
- Identity + owner
- Model, version, release, and accountable team
- Intended users
- Who may operate or integrate it
- Supported tasks
- Uses and conditions covered by evidence
- Unsuitable uses
- Unsupported, unsafe, or prohibited contexts
- Training context
- Data description, objective, and known unknowns
- Evaluation
- Methods, metrics, samples, and segmented results
- Limitations
- Known failures, uncertainty, and mitigation dependencies
- Operations
- Monitoring, updates, reporting, and retirement
A model card is a transparency artifact—not certification—and becomes unsafe when self-reported claims, test context, or version information go stale.
02Why it matters
Teams buying or adapting a model need evidence beyond a leaderboard score. A card helps expose mismatches between a supplier's tests and the proposed product, identify missing evaluation, and preserve assumptions when ownership changes.
03When to use it
Read or create a model card when selecting, releasing, adapting, or materially updating a model. Scale detail to the decisions readers must make. For a product composed of several models and tools, supplement the cards with evaluation of the assembled system.
04How it works
- Identify the model version, owner, intended users, supported tasks, and deployment assumptions.
- Describe training and evaluation data at an appropriate level, linking dataset documentation and recording unknowns.
- Report evaluation methods, metrics, sample sizes, and results across relevant groups and operating conditions.
- Record unsuitable uses, known failures, mitigation dependencies, and unresolved questions without presenting them as solved.
- Review the card with product and technical owners, publish to the appropriate audience, and update it with model releases.
05A practical example
A support platform considers a sentiment classifier whose card reports tests only on English consumer reviews. The PM identifies a mismatch with multilingual business support tickets and requires local evaluation before using its scores to prioritize queues.
06Where it breaks
Cards are often self-reported, incomplete, or stale. Results depend on the test population and cannot establish performance in a new workflow. Some training details may be unavailable, and publishing a limitation does not control its consequences.
Common mistakes
- Publishing only favorable aggregate scores without evaluation conditions.
- Treating a supplier's model card as approval for the entire product.
- Copying documentation across versions without checking changed behavior.
- Using intended-use statements to imply unsupported guarantees.
07Sources
- Model Cards for Model ReportingACM Conference on Fairness, Accountability, and Transparency · Peer-reviewed researchModel Cards for Model Reporting
- Model CardsHugging Face · Official documentationModel Cards