How it works
Created by companies and researchers, a model card includes information such as architecture, training data, performance metrics, known biases, limitations and usage instructions. It is a transparency standard adopted by Hugging Face, Google, Meta and others.
The concept was proposed by Google researchers in 2019. The idea is that every AI model should have a "card" that helps users understand what the model does, what it is for and what its risks are.
In practice, model cards vary in quality. Some are detailed and useful. Others are generic and do not provide enough information to make informed decisions.
Practical example
Before using an open source model like Llama 3, a developer can check the model card on Hugging Face to understand which languages it was trained on, which tasks it performs well on and what bias risks exist.
The Llama 3 model card includes information about training data, benchmark metrics, known limitations and responsible use instructions. This helps the developer decide if the model is suitable for their application.
What it's for
Make informed decisions about which model to use, identify risks and ensure responsible use. For companies, model cards help assess compliance with internal policies and regulations.
For developers, it is the first stop before adopting a model. Reading the model card can avoid surprises in production, such as unexpected bias or language limitations.
Difference between model card and README
README is the general repository documentation (how to install, how to use). Model card is focused on the model itself (characteristics, limitations, risks). A repository can have both. The model card is more standardized and focused on transparency.
In practice, many repositories on Hugging Face use the model card as a README. But the concept is distinct: README is technical documentation, model card is responsibility documentation.