A model’s architecture is its underlying structure. Information about a model’s architecture can provide insight on how the model was trained and developed.

For example, a model might be a transformer-based model, as used by many large language models, or a diffusion model of the kind used in some image generation apps.

This is a description of the inputs and outputs of a model.

Inputs for models can include many modalities, including text, images, audio, video, and more, depending on the model.

Similarly, a model’s outputs may be a simple text-based answer, images, audio, videos, a combination of these, and much more.

Some model cards will include information on the model’s terms of use, with links to any further technical documentation developers might find useful.

This section of a model card will generally include information about the types of data used to train a model, as well as any methods and processes used to clean and filter training data to adhere to safety policies.

Intended usage describes what the model is designed to do, and how it works for end users. This entry may include any specific applications or platforms the model has been developed for.

The information here is valuable because it helps developers understand what the model was designed to do and the benefits of using this model relative to others.

This entry describes the potential known limitations of the model.

Some model cards will address factors that might limit the performance of the model and what conditions must be satisfied for the model to be used effectively.

This entry provides insights into the type of hardware used in the model’s development and training, and why that hardware was chosen.

If the model depends on any specific hardware for its ongoing use, it may also be listed here.

This entry describes the software that was used to train the model, and may include details on the benefits of using the software during the model’s training in comparison to other software and approaches.

This entry provides information about the types of evaluations used to test the model's performance.

The methodology for the evaluations may also be explained so the results can be understood and put into a useful context.

Any industry and academic benchmarks used to evaluate the model may also be included.

This section of a model card explains the approach used to evaluate the model for any responsibility concerns.

Any specific safety testing techniques used such as adversarial testing and red teaming will be listed here where relevant.

The findings of any evaluations, potential risks of the model, and the mitigations used to address any concerns, may also be listed here.