This is a template for a model card for model reporting.
The template is here.
Model cards for model reporting was created to increase transparency of models.
[Model cards are] short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups ... and intersectional groups ... that are relevant to the intended application domains. Model cards also disclose the context in which models are intended to be used, details of the performance evaluation procedures, and other relevant information.}
The problem it is trying to solve:
Currently, there are no standardized documentation procedures to communicate the performance characteristics of trained machine learning (ML) and artificial intelligence (AI) models. This lack of documentation is especially problematic when models are used in applications that have serious impacts on people’s lives such as in health care, employment, education and law enforcement.
One of the main concerns of model cards is ethics. As such, it emphasizes metrics across intersectional groups. That is, performance not only in the obvious, larger groups, such as male vs. female, but also in combinations of those groups, "male, Fitzpatrick skin type I" vs. "male, Fitzpatrick skin type V", or "female, ages 18-34" vs. "female, ages 35-50". Disaggregating the measures of a model before putting it in production can prevent embarrassing and potentially harmful errors, such as the very public shaming of Microsoft and IBM in the Gender Shades paper and accompanying MIT Media Lab's website.
The short explanation: using a markdown file allows us to compare (diff) easily one version of the model card with another version.
The longer explanation:
Models should be under version control, in the same way we put the code under version control. Once under version control, we can compare one version against the other.
It is easier to follow the changes in a model when its model card is distributed together with the model. If the model is under source control, so should be its model card.
Whenever there is a new version of the model, we also need to update its description. In other words, we need to update its model card.
The model card distributed with a version should be in a format that is easy to compare with previous versions, to allow us to quickly see what has been changed. Markdown is a simple, text format, making it ideal for that.
- CheXNet
- Google's interactive model cards for face detection and object detection
- NVIDIA's DashCamNet
- OpenAI's GPT-3 and CLIP
- Hugging Face has a place for model cards. See this example from the bert-uncase-model.
- Amazon introduced AI Service Cards late in 2022. An example: Rekognition Face Matching service card.
- PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models added a model card to its paper after discussions about biases in the model.
If you are interested in model cards, you may also want to review datasheets for datasets.