Model Card

A model card is a short document that ships with a machine-learning model to describe its intended use, training data, performance across demographic groups, and known limitations. It gives downstream teams the context needed to decide whether a model fits a task.

Researchers Margaret Mitchell and colleagues at Google proposed the format in a 2019 FAT* conference paper. The idea moved from academic proposal to standard practice, and platforms like Hugging Face now render every uploaded model’s README as a model card.

Every card follows a template: model details, intended use, factors, metrics, evaluation and training data, subgroup performance, ethical concerns, and caveats. Buyers and regulators lean on that structure to compare vendors like a procurement checklist.

Regulators pushed the format from best practice into law. The EU AI Act, which entered into force in August 2024, requires providers of general-purpose AI models to publish public technical documentation that maps closely onto the sections Mitchell proposed.

Key takeaways

• A model card documents an ML model’s intended use, training data, performance by subgroup, and known limitations in one short file.
• The format was proposed by Margaret Mitchell and coauthors at Google in a 2019 paper and now anchors Hugging Face model repos.
• Sections cover model details, intended use, factors, metrics, evaluation data, training data, quantitative analysis, ethics, and caveats.
• Cards are effectively mandatory under the EU AI Act’s transparency provisions for general-purpose AI models placed on the market.
• Buyers use cards to shortlist vendors; auditors use them to check bias, safety, and documentation gaps before deployment.

How it works

A model card compresses a model’s operating envelope into a short structured document. Sections run from model details to caveats, each answering a specific question a downstream user or auditor tends to ask.

The template is deliberately structural. Each section forces a specific piece of evidence into the open — training data provenance, subgroup breakdowns, ethical risks — so gaps become visible rather than invisible.

If a team cannot fill a row, that empty row is itself a finding. A missing subgroup evaluation, for instance, forces the reviewer to ask why — and answers to those questions often catch shipping risks before launch.

Teams publish the card in the model repo. Hugging Face, for example, renders a repo’s README as its card and parses the YAML metadata block at the top into filter facets. Data scientists write it; legal, security, and buyers read it.

Examples

Model cards now ship with almost every major foundation model, and their tone has shifted from research curiosity to compliance artifact. Below are four cards that shaped how the industry writes them today.

Google’s Face Detection card (2019) was Mitchell and colleagues’ reference implementation. It reported false-positive and false-negative rates across skin type, age, and gender, and flagged the model as unsuitable for surveillance.

Meta’s Llama 2 (July 2023) shipped model cards for seven variants, from 7B to 70B parameters, alongside a responsible-use guide covering safety fine-tuning, red-team results, and prohibited applications. Each card links to Meta’s 77-page technical report.

OpenAI’s GPT-4 system card (March 2023) popularised the term system card, a broader sibling covering deployment risks like disinformation, weapons uplift, and privileged access. It shipped as a 60-page annex to the GPT-4 technical report.

Hugging Face enforces model cards at the platform level. Every uploaded model repo renders its README.md as a model card, and YAML metadata at the top feeds discoverability filters for license, language, task, base model, and CO2 emissions.

Related terms

Model cards sit inside a broader family of AI transparency artifacts, some documenting the model itself, others the data or the deployed system. Knowing which is which prevents shipping the wrong document to the wrong audience.

• Artificial Intelligence: umbrella field for systems that mimic human reasoning, learning, and decision-making across data and tasks.
• Machine Learning: the subset of AI in which models learn statistical patterns from labeled or unlabeled training data.
• Generative AI: models that produce new text, image, audio, or code output rather than classify inputs.
• Natural Language Processing: the branch of AI concerned with parsing, understanding, and producing human language at scale.
• Data Science: the discipline that trains, evaluates, and interprets the models a card describes.
• Automation: the practice of routing routine, rule-based work to software agents guided by documented models.

FAQ

Who writes a model card?

The model’s development team writes the first draft, usually the data scientists and ML engineers who trained it. Legal, safety, and product review the card before release. Cards get updated whenever the model is retrained or fine-tuned.

Are model cards legally required?

The EU AI Act, which entered into force in August 2024, requires providers of general-purpose AI models to publish documentation that maps closely to the model-card template. In the US, the NIST AI Risk Management Framework recommends the same shape.

What is the difference between a model card and a system card?

A model card documents a single model — its data, metrics, and limits. A system card zooms out to the whole product, covering guardrails, deployment surfaces, and downstream risks such as misuse or jailbreak paths.

How long should a model card be?

Cards range from a page to sixty. Narrow classifiers stay short, while foundation-model cards run long because they cover training data, safety tests, and multiple downstream tasks in one document. Anthropic and OpenAI cards regularly clear 40 pages.

Where do outsourced teams fit into model-card work?

Outsourced annotation and QA vendors often supply the raw evidence that fills the training-data and quantitative-analysis sections, from labeled samples to bias-test results and red-team logs.

Outsource Accelerator connects AI teams with specialist BPOs that handle model-card evidence work like annotation, evaluation, and red-team support.