Model Card

A model card is a standardized document that discloses the intended use, performance, and limitations of a machine learning model.

Model cards were introduced by Mitchell et al. at Google in 2018 as a framework for transparent reporting of machine learning models. They are designed to accompany a model when it is released, providing essential information to users and stakeholders. The document typically includes sections on model details, intended use, factors affecting performance, evaluation metrics, training data, ethical considerations, and caveats.

The purpose of a model card is to reduce information asymmetry between model developers and end users. By clearly stating what a model was designed for, what data it was trained on, and where it might fail, model cards help prevent misuse and promote responsible deployment. They are analogous to nutrition labels for food, offering a concise summary of a model’s characteristics and potential risks.

Model cards are part of a broader movement toward responsible AI practices. They complement other documentation tools like datasheets for datasets and system cards. While not universally adopted, they have been influential in shaping industry standards for model transparency and are increasingly required by regulatory frameworks and organizational policies.

Why it matters

Model cards matter because they operationalize transparency in machine learning. Without them, users may deploy models in inappropriate contexts, leading to unfair outcomes or safety risks. By providing clear, structured information, model cards empower practitioners to make informed decisions about model selection, evaluation, and monitoring, ultimately fostering trust and accountability in AI systems.

First appeared

Mitchell et al., Google, 2018.

FAQ

How does it work?

A model card is created by the model developer and includes predefined sections that describe the model’s purpose, performance metrics, training data, and limitations. It is typically published alongside the model, allowing users to quickly assess whether the model is suitable for their task and to understand potential biases or failure modes.

What information is typically included in a model card?

Common sections include model details (architecture, version), intended use (primary and secondary uses), factors (demographic or environmental conditions that affect performance), evaluation metrics (accuracy, fairness metrics), training data (source, size, preprocessing), ethical considerations, and caveats (known limitations or recommended usage restrictions).

How does a model card differ from a datasheet?

A model card focuses on the model itself—its design, performance, and intended use—while a datasheet documents the dataset used for training or evaluation. Both aim to increase transparency, but they address different components of the machine learning pipeline. Model cards and datasheets are often used together for comprehensive documentation.