Cloud Hosting

What is a model card in machine learning?

A model card is a type of documentation that is created for, and provided with, machine learning models. A model card functions as a type of data sheet, similar in principle to the consumer safety labels, food nutritional labels, a material safety data sheet or product spec sheets.

There has been a dramatic rise in the development and adoption of machine learning (ML) and artificial intelligence (AI) during recent years. Further advances in generative AI employ large language models (LLMs) as a core component. However, the many models used in those platforms are increasingly complex and difficult to understand. Even model developers sometimes struggle to fully understand and describe the ways a given model behaves. This complexity has created serious questions about core business values such as transparency, ethics and accountability. Common questions include the following:

First proposed by Google in 2018, the model card is a means of documenting vital elements of a ML model so users -- including AI designers, business leaders and ML end users -- can readily understand the intended use cases, characteristics, behaviors, ethical considerations, and the biases and limitations of a particular ML model.

As late as 2024, there are no current legislative or regulatory requirements to produce or provide model card documentation with ML models. Similarly, there are no currently established standards in model card format or content. However, major ML developers have spearheaded the adoption of model card documentation as a way of demonstrating responsible AI development, and adopters can find model cards for major platforms such as Meta Llama, Google face detection and OpenAI GPT-3.

The rise of ML and AI is driving the need for transparency and responsible governance. Businesses must understand what ML models are for, how they work, how they compare to other competitive models, how they're trained and their suitability for intended tasks.

Model cards are a tool that can address such concerns which readily impact governance and regulatory issues for the business. Model cards can provide a range of important benefits to ML and AI projects including the following:

Labels and other informational summaries are generally most effective when they allow comparing similar products side by side using comparable content and formats. However, the information presented on an ML model card can vary. Unlike highly regulated informational displays -- such as food nutritional labeling -- there are no current standards to govern the information or formatting included on ML model cards.

ML models can vary dramatically in their scope, purpose and capabilities, so this makes it hard to regulate. For example, an ML model developed to aid in medical diagnosis can be distinctly different than an ML model created to run analytics on retail sales operations, or a complex LLM used in an AI construct. Consequently, ML model developers largely use their own discretion to determine what information to include, and how that information should be presented. Yet, as leading technology firms develop ML/AI platforms and document those offerings through model cards, some de facto documentation standards are taking shape. Model cards should include the following:

This first section of a model card is typically the introduction to the model which can outline the model's essential details including the model's name, version, revision list, a brief general description of the model, business or developer details and contact information, and licensing details or limits.

This section describes the intended uses, use cases and users for the model. For example, a section on use cases may describe uses in object detection, facial detection or medical diagnoses. This section may also include caveats, use limitations or uses deemed out of scope. For example, a model intended for object detection may detail input from photos or video; output including detection of a specified number of object classes; and other output data such as object bounding box coordinates, knowledge graph ID, object description and confidence score.

This section describes the overall design of the model and any underlying hardware back end that runs the model and hosts related data. Readers can refer to the model card to understand the design elements or underlying technologies that make the model work. For the object detection model example, the model card may describe an architecture including a single image detector model with a Resnet 101 backbone and a feature pyramid network feature map.

This section outlines, describes or summarizes the data used in model training; where and when the data was obtained; and any statistical distribution of key factors in the data which may allow for inadvertent bias. Since training data may be proprietary to the model's developers, training details may be deliberately limited or protected by a separate confidentiality agreement. Training details may also describe training methodologies employed with the model.

This section outlines details related to the model's performance measured against a test data set, not a training data set, as well as details about the test data set itself. For the object detection model example, performance metrics included on the model card may note the use of both Google's internal image data set as well as an open source image set as test data and the number of object classes the model can detect in each data set. Additionally, performance details may outline reported metrics including the precision and accuracy of the object detection. More sophisticated models may utilize other detailed metrics to measure performance.

A key segment of any model card is the section describing limitations, possible biases or variable factors that might affect the model's performance or output. For the object detection model example, known limitations may include factors such as object size, clutter, lighting, blur, resolution and object type since the model can't recognize everything.

This final segment of a model card is often dedicated to business-related details including information about the model's developers, detailed contact, support and licensing information, fairness/privacy and usage information, suggestions for model monitoring, any relevant assessment of impacts to individuals or society, and other ethical or potential legal concerns related to the model's usage.

As leading technology organizations build ML and AI platforms, their work on model cards and other documentation has provided a standard for other ML firms to follow. Today there are many examples of ML model cards to review including the following major examples:

There are also more standardized tools for model card creation, as well as model card repositories, such as these examples:

Both GitHub and Hugging Face provide a repository of model cards which are available for review and study, offering model card examples across many different model types, purposes and industry segments.

Original post:
What is a model card in machine learning and what is its purpose? - TechTarget