Model Card: Definition and Examples

A model card is a reference document that provides essential information about an artificial intelligence model. Introduced by Google researchers in 2019, it aims to improve transparency and accountability in the development and deployment of AI systems. It plays a role comparable to a medication insert: it informs the user of what the model does, how it was trained, and under what conditions it works correctly.

A typical model card contains several key sections: a general description of the model, training data used, performance metrics evaluated on different population subgroups, known limitations, identified biases, and usage recommendations. It may also include information about the carbon footprint of training, ethical considerations, and use cases for which the model is not suitable.

In prompt engineering, consulting a model card is a fundamental step before designing prompts. It allows you to understand the model's strengths and weaknesses, identify areas where it excels and those where it risks producing unreliable results. For example, if a model card indicates that the model was primarily trained on English data, this will guide prompting strategy for multilingual tasks.

Today, model cards have become standard practice in the industry. Platforms like Hugging Face integrate them directly into their model repositories, and organizations like Anthropic, OpenAI, and Meta publish detailed technical documents (often called system cards or technical reports) that serve a similar function for their advanced models.