Leveraging Documentation: Implementing ML Model Cards for Better Decision Making

AI offers enormous potential for businesses, individuals or countries, but also comes with risks on the ethical and societal level. These include failed or flawed systems, biased outcomes, or the misuse of algorithms for unintended purposes.

Transparency in the development and deployment of AI systems, as well as model and data documentation play an important role to ensure responsible use of AI. Documentation and availability of information becomes especially important regarding applications with severe impact on people’s lives (e.g. finance, employment, healthcare, or education) as they are subject to regulation, e.g. in the upcoming EU AI Act.

But you shouldn’t only care about a good documentation because of regulation or responsible AI. It actually helps you foster collaboration and knowledge sharing within and between teams.

Besides all the benefits, documentation is often neglected due to the manual overhead it can cause. One way to reduce manual effort while increasing transparency for all stakeholders are “Model Cards”. Keep reading to find out what they are, how you can implement them & how they can help you reduce documentation time by a factor of 10!

What are Model Cards?

The term “Model Cards” was introduced in a research paper by Google (”Model Cards for Model Reporting” by Mitchell et al., 2018). It refers to a short document containing the most important information about a machine learning (ML) model. The paper proposes a framework for standardized reporting and documentation of ML models, with the aim of promoting transparency, fairness, and responsible deployment in real-world applications.

Model Cards as Part of Model Documentation

When developing ML algorithms and applications it is crucial to document important aspects of the development to ensure reproducibility, transparency and appropriate use.

A thorough documentation of the ML development process consists of a code, data, and model documentation incl. the training process. Model Cards provide you a template/structure about what to add to your documentation. Automating the creation of Model Cards, can then lead to improvements both in quality and speed of documentation.

Model Cards provide insights into the model's intended use, training data, evaluation protocols, performance metrics, and limitations / considerations related to fairness and potential societal impacts.

Clear and concise model documentation enables stakeholders to understand the model's behavior and make informed decisions regarding its deployment and usage.

Components of a Model Card

The original framework by Google suggests nine factors that build up a Model Card, though organizations can modify this construct according to their preferences. Their aim is to provide holistic information about key factors that can influence the behavior and performance of machine learning models.

1. Model Details: This section provides an overview of the model, including its name, version, and purpose. It may also include details about the model's architecture, parameters, and any pre-trained components used. It doesn’t need to contain sensitive information the organization doesn’t want to share, but rather convey basic information.
2. Intended Use: Model Cards describe the intended use cases and applications of the model, specifying the tasks it is designed to perform and the target audience. Quickly conveying why, for whom and for what purpose the model was created.
3. Factors: A summary of model performance across a variety of relevant factors including groups (e.g. people of the same gender), instrumentation (e.g. which camera was used for taking the input pictures), and environments (e.g. lightening conditions the model is deployed in). These external factors are included to point to potential limitations of the model.
4. Metrics: List of performance metrics and decision thresholds. Be aware that suitable metrics vary with the type of model tested.
5. Evaluation Data: Provide information about the source and composition for all referenced datasets used for validation and testing.
6. Training Data: Ideally follows the same level of transparency as evaluation data, but often unfeasible due to privacy restraints of the organizations. For external communication basic details about the distributions over groups in the data & potential biases should suffice.
7. Quantitative Analysis: Results of evaluating the model according to the chosen metrics, broken down by relevant factors.
8. Ethical Considerations: This part discusses potential bias, fairness, and ethical considerations related to the model and its deployment. It may address issues such as algorithmic bias, fairness in prediction outcomes, and steps taken to mitigate these concerns.
9. Caveats & Recommendations: Place for additional concerns not covered in the previous sections (awareness information when using the model, limitations, risks, dependencies or trade-offs).

3 Reasons to start using Model Cards

Make Better Informed Decisions

Model Cards enable data scientists and decision-makers to gain deeper insights into a model's performance and limitations. Developers can compare the model’s results to other models in the same space to learn from the past and save time experimenting and training. ML and AI practitioners can better understand how well the model might work for the intended use cases and track its performance over time.

Furthermore, decision-makers can assess risks associated with model deployment and inform product design decisions based on the model's performance characteristics.

Decrease Documentation Time

Manual documentation can be time-consuming and is often outdated. By automating this process, Model Cards significantly reduce the time spent on documentation. This streamlines the documentation process for data scientists and alleviates the burden of writing extensive documentation, ultimately saving valuable time that can be used for developing.

The Data Science Team at Wayflyer built their own internal tool, following the concepts of Model Cards using various open source solutions. In an interview their Data Science Manager told us his team was “able to decrease the time spent on documentation by a factor of ~10 with automated Model Cards”.

Increase Transparency and Knowledge Sharing for all Stakeholders

Model Cards enhance transparency by providing standardized documentation, fostering accountability, and promoting ethical considerations. They facilitate knowledge sharing and collaboration across the organization and the wider community, like customers or policy-makers. By sharing insights, organizations can build upon existing models, bridge the gap between business and tech, and empower stakeholders to understand the model's behavior, limitations, and potential biases.

By adopting Model Cards, organizations can make more informed decisions, streamline documentation processes, and enhance transparency and collaboration, ultimately driving responsible and efficient use of machine learning models.

How to implement Model Cards at your Organization

Now is the best time to get started on implementing Model Cards in your organization to leverage above benefits for yourself and prepare for upcoming regulation.

You can develop a custom solution, leverage and combine open source frameworks, or trust a plug & play solution.

Integrate Model Cards using Open Source Solutions

The team behind the original research paper also published an open source Google Model Card Toolkit on github. The full functionality of the tool only works with TensorFlow, limiting the functionality for the broad mass.

The team from Wayflyer had this problem and started building their own internal tool, making use of open source tools like Pydantic and Jinja and taking inspiration from the original Model Cards framework. The output are basic HTML files. Learn more about how Wayflyer did it in this article.

One Line of Code away from Automated Model Cards: trail

If you want to get started quickly and still customize the Model Cards to your needs, you should take a look at trail.

trail easily integrates in your development environment and tracks and aggregates all information from metrics, to parameters, and meta information from your code. On top of creating Model Cards for every experiment, trail also helps with bringing those experiments into context and stores the reasoning behind the process in a tree-like format.

trail automizes your Model Card documentation and makes sure model, data and enriching information are standardized and easy to digest.

Additionally Google’s Model Cards face the problem that they are often very technical and hard to digest for non-technical people, even though this is what Model Cards should facilitate. trail leverages AI to prepare different abstraction levels of documentation to ensure, everyone understands the use cases, implications and limitations before deploying a model.

Don’t wait any longer and start implementing the system of Model Cards today with the help of trail. Get early access here.