Model card

Describing a model’s conditions of use clearly

A model card is documentation accompanying an AI model. It describes matters such as intended use, training and evaluation data, evaluation methods and known limitations. The information should be associated with an identifiable model version. Results become meaningful alongside their evaluation conditions: which tasks, inputs and languages were examined? The card supports an informed choice but does not replace assessment in the intended local use case.

For a fictional model that categorises service reports, a card would name the type of report considered and the categories assessed. If quality had been evaluated only on standardised written reports, this would not establish assessed suitability for automatically transcribed voice messages containing workplace abbreviations. That input type would need to remain visible as an open question and be examined separately. The information would need checking again for a new model version. A clearly written card is not, by itself, an independent certificate or evidence of error-free routing.

Sources

• Hugging Face · Model Cards

• Mitchell et al. · Model Cards for Model Reporting