Model Card
A model card is a structured, standardized document accompanying a machine learning model that discloses its intended use, training data, performance metrics, limitations, and known risks, intended to help downstream users evaluate whether…
Definition
A model card is a structured, standardized document accompanying a machine learning model that discloses its intended use, training data, performance metrics, limitations, and known risks, intended to help downstream users evaluate whether the model is appropriate for their use case.
Overview
As machine learning models, and especially large pretrained models, have become widely reused across many downstream applications by people other than their original creators, the need for clear, consistent documentation about what a model does and does not do well became apparent. Model cards were formally proposed in a 2019 paper, 'Model Cards for Model Reporting,' by researchers including Margaret Mitchell and Timnit Gebru while at Google, drawing an analogy to nutrition labels or datasheets that accompany consumer products and hardware components. A typical model card covers: the model's intended use cases and out-of-scope uses; details of the training data and any known biases or gaps in it; the model architecture and training procedure; quantitative performance metrics, ideally broken down across relevant subgroups (e.g., by demographic, language, or domain) rather than a single aggregate number, since aggregate metrics can mask disparities in performance; known limitations, failure modes, and ethical considerations; and sometimes environmental impact figures such as training compute and estimated carbon footprint. The goal is to give anyone deciding whether to use, fine-tune, or build on top of a model enough information to make an informed judgment, rather than treating the model as an unexamined black box. Model cards have since become a widely adopted convention, particularly on model-hosting platforms like Hugging Face, where nearly every uploaded model includes a README-style model card, and among major AI labs, some of which publish detailed 'system cards' for large frontier models (a related but often more extensive document covering safety evaluations, red-teaming results, and deployment considerations for a specific model release). While adoption and thoroughness vary widely — many community-uploaded model cards are sparse or incomplete — the practice is now considered a baseline expectation for responsible model release, and is referenced in various emerging AI governance and regulatory frameworks as a documentation standard.
Key Concepts
- Standardized document describing a model's intended use, training data, and limitations
- Includes performance metrics broken down by relevant subgroups, not just aggregate scores
- Discloses known biases, failure modes, and ethical considerations
- Originated from a 2019 Google research paper drawing an analogy to product nutrition labels
- Widely adopted on model-hosting platforms such as Hugging Face
- Related to, but generally shorter than, 'system cards' published for frontier AI model releases
- Helps downstream users judge whether a model is fit for their specific use case
- Increasingly referenced in AI governance and regulatory documentation requirements