Model Card: What It Is, Examples, Templates, How to Create

A model card is a short document that explains how a machine learning model works, what it was built for, and what its limitations are. Think of it as a nutrition label for AI. It tells you what went into the model, how it performs, and where it might fail.

Model cards emerged in 2018 as a way to bring transparency to AI systems. They help developers understand whether a model fits their needs, help regulators assess risks, and help organisations decide whether to adopt a particular AI tool. When you build or buy AI systems, model cards give you the facts you need to make informed decisions.

This guide explains what goes into a model card, why they matter for your AI projects, and how to create one. You’ll see real examples from organisations like Meta and Google, find templates you can use, and learn how model cards fit into enterprise AI governance. Whether you’re evaluating a vendor’s model or documenting your own, you’ll understand what makes a model card useful and how to put one together properly.

Why model cards matter in AI projects

Most AI projects fail not because of bad models, but because teams don’t understand what they’re deploying. Your organisation might spend months training a model only to discover it performs poorly on your specific data, or worse, produces biased results that damage your reputation. Model cards solve this problem by making AI systems transparent before you commit resources to implementation.

They reduce risk in AI adoption

When you evaluate a model card, you see exactly what data trained the model, which means you can spot potential biases before they affect your customers. A model trained primarily on data from one demographic group might fail spectacularly when applied to a different population. You also learn about the model’s performance limits, so you don’t deploy a system for use cases it wasn’t built to handle. This visibility prevents costly mistakes and helps you set realistic expectations with stakeholders.

Model cards help you avoid the trap of assuming a model that works well in one context will automatically work in yours.

They support governance and accountability

Your compliance team needs documentation to prove your AI systems meet regulatory requirements. Model cards provide structured evidence of responsible AI practices, showing regulators and auditors that you understand what your models do and don’t do. They also establish clear ownership by documenting who built the model, when, and for what purpose. When something goes wrong, you have a record that traces decisions back to specific choices about training data, architecture, and intended use.

How to create a model card

Creating a model card doesn’t require weeks of effort or complex tools. You start by gathering information about your model as you build it, then organise that information into a structured document. The process becomes easier when you document decisions in real time rather than trying to reconstruct everything after deployment. Most teams complete a thorough model card in a few hours once they have the underlying data about their model’s development and testing.

Start with the basics

Begin by identifying the fundamental facts about your model. You need to know who built it, when it was created, what version you’re documenting, and what problem it solves. Write down the intended use cases clearly, because this section prevents misuse later. If you built a model to detect manufacturing defects in one type of product, state that explicitly rather than describing it vaguely as a “quality control system”. This specificity helps future teams understand whether your model fits their needs.

Document your decisions as you build

Track your choices throughout development rather than waiting until the end. Record which training datasets you used, why you chose them, and what preprocessing steps you applied. Note the model architecture and any constraints you implemented for fairness or performance. When you run evaluation tests, save the results immediately along with details about the test conditions. This ongoing documentation makes model card creation straightforward because you’re not relying on memory or trying to reverse-engineer your own work.

The best model cards come from teams that treat documentation as part of development, not an afterthought.

Use existing frameworks and tools

Several organisations provide templates and tools that structure your model card automatically. Google’s Model Card Toolkit generates cards from your model metadata and evaluation results. You can also start with the original research template that defines standard sections and adapt it to your needs. These frameworks ensure you don’t miss critical information and make your model cards consistent across projects. Choose a template that matches your industry’s requirements, then customise it based on your stakeholders’ specific concerns. Your legal team might need more detail about data provenance, while your product team might care more about performance metrics across different user segments.

What to include in a model card

Your model card needs specific sections that give readers a complete picture of how your AI system works and where it might fail. These sections follow a standard structure that researchers established in 2018, though you can adapt them based on your organisation’s needs. Each section answers critical questions that developers, compliance teams, and decision-makers ask before deploying AI systems.

Core identification sections

Start with model details that cover the basics: who built the model, when, what version you’re documenting, and what type of architecture you used. You should include contact information for questions and link to any research papers or documentation. Next, define the intended use clearly by listing specific applications the model was designed for and explicitly stating what falls outside its scope. A model built to detect fraud in credit card transactions, for example, shouldn’t be repurposed for loan approval decisions without additional validation.

Performance and data sections

Document your training data by describing where it came from, how you collected it, and what preprocessing steps you applied. You don’t need to expose proprietary methods, but you must provide enough detail for readers to understand potential biases or limitations in the data. Include performance metrics that show how the model performs across different conditions, demographic groups, or environmental factors. Test your model on various scenarios and report the results honestly, including where it underperforms. This transparency helps teams decide whether your model suits their specific context.

Model cards that hide weak performance in certain conditions create more problems than they solve.

Governance sections

Your model card needs a quantitative analysis section that presents evaluation results in detail, showing accuracy, precision, recall, and other relevant metrics. Break these down by meaningful categories like user demographics or usage conditions. Add an ethical considerations section that addresses fairness concerns, privacy implications, and potential societal impacts from the model’s use. Finally, include recommendations for monitoring, testing, and validating the model in new contexts. These sections demonstrate responsible AI practices and help your compliance team prove you’ve thought through the risks before deployment.

Model card examples and templates

You learn best by seeing how successful organisations document their AI systems. Several major technology companies publish model cards that demonstrate different approaches, each tailored to their specific needs and audiences. These examples show you what works in practice rather than just theory, and you can adapt their structures for your own models. Templates save you time by providing a starting framework that ensures you cover essential information without reinventing the structure.

Real-world examples from major organisations

Meta and Microsoft’s Llama 2 model card sits within their research paper’s appendix and includes detailed sections on hardware requirements, carbon footprint, and ethical considerations. This approach works well when your technical audience needs deep implementation details alongside governance information. OpenAI’s GPT-3 model card takes a simpler approach with clear sections on model details, intended use, data sources, and limitations, plus a direct link for submitting feedback. Google’s face detection model card stands out because it uses visual examples to illustrate limitations, showing images of blurry faces or crowded scenes where the model struggles. This visual approach helps non-technical stakeholders understand constraints quickly.

The best model cards match their format to their audience rather than following a rigid template.

Where to find templates

Google’s Model Card Toolkit provides a structured framework that generates model cards from your metadata automatically. You input your model details, training data information, and evaluation results, then the toolkit formats everything into a standardised document. Amazon SageMaker Model Cards offer another template integrated directly into their cloud platform, making documentation part of your development workflow. Hugging Face model cards show how open-source communities structure their documentation, focusing on practical implementation details that developers need. Start with whichever template matches your development environment, then customise sections based on your stakeholders’ specific concerns about fairness, performance, or compliance.

Using model cards in enterprise AI

Your enterprise needs consistent documentation across all AI systems, whether you build them internally or buy them from vendors. Model cards provide that standardisation by giving your teams a single format for evaluating, comparing, and managing AI models at scale. They turn abstract concepts like “responsible AI” into concrete checkpoints that your organisation can verify before deploying any model into production. This standardisation becomes critical when you manage dozens or hundreds of AI systems across different departments.

Evaluating third-party AI systems

When vendors pitch their AI solutions, you need to ask for their model card before signing any contracts. A comprehensive model card reveals whether the vendor trained their model on data similar to yours, tested it across relevant user groups, and documented known limitations. You can compare multiple vendors’ models side by side using their model cards, focusing on metrics that matter to your business like accuracy across demographics or performance in specific conditions. Vendors who refuse to provide detailed model cards raise immediate red flags about transparency and accountability.

Model cards transform vendor selection from trust-based decisions into evidence-based evaluations.

Establishing internal governance frameworks

Your organisation should require model cards for every AI system your teams build. Create a standard template that includes sections your legal, compliance, and product teams need to review before deployment. Set clear expectations that documentation happens during development, not afterwards, so teams capture decisions about training data, architecture choices, and testing results in real time. This process ensures your internal models meet the same transparency standards you demand from external vendors.

Final thoughts on model cards

Model cards give you the transparency and accountability your AI projects need to succeed. They transform vague promises about responsible AI into documented evidence that your stakeholders can review and trust. Your organisation benefits most when you treat model cards as essential documentation rather than compliance paperwork, building them into your development workflow from day one.

Start documenting your AI systems now, whether you build models internally or evaluate vendor solutions. Every model card you create reduces risk, improves decision-making, and strengthens your governance framework. If you need help establishing model card standards across your AI initiatives or want guidance on preparing your data for AI implementation, get in touch with our team. We help organisations build transparent, production-ready AI systems that deliver lasting business value.