Model card
A short document attached to a model. It states intended use, populations evaluated and known limits, so that out-of-domain use becomes a demonstrable fault.
A model card is a standardized document attached to a published model. It sets out intended and discouraged uses, training and evaluation data, performance broken down by subpopulation rather than averaged, and known limits. The founding idea is simple and carries far. A model's average performance hides considerable gaps between groups; a disaggregated report shows what a single metric conceals. The legal significance goes further than the engineering one. By declaring a domain of use, the provider makes any use outside that domain the deployer's act, and in turn binds itself to what it declared. The card becomes the document around which responsibility is split between whoever builds and whoever applies. Recent regulatory regimes have adopted that split by requiring technical documentation of the same kind.
The model card was proposed in the paper Model Cards for Model Reporting, presented in January 2019 at the conference on fairness, accountability and transparency. It has since become the usual documentation format for publicly released models.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.