Hugging Face Model Card Guidelines

These guidelines tell anyone publishing a model on Hugging Face what they must include in their model card and how to set key metadata fields. You are required to document your model's intended uses, limitations, biases, training data, training details, and evaluation results. Metadata like the license, pipeline tag, and library name are not just labels—they control what gets displayed on your model page and how other users can find and interact with your model.

The Hugging Face Model Card Guidelines establish mandatory disclosure and metadata requirements for models hosted on the Hugging Face Hub. Creators are required to include in the model card: intended uses and potential limitations (including biases and ethical considerations), training datasets, training parameters and experimental information, and evaluation results. Metadata fields—including license (with custom licenses requiring an 'other' value plus a name and link), pipeline tag, and library_name—carry direct functional consequences: the pipeline tag controls the interactive widget and underlying APIs, structured evaluation results are parsed and rendered as a Hub widget, and the license is publicly displayed and used as a filter. For model repos created after August 2024, explicit declaration of library_name: transformers is required, as automatic detection no longer applies.

For a user browsing or downloading models on Hugging Face, these guidelines mean that model pages are required to disclose intended uses, limitations, biases, ethical considerations, training datasets, and evaluation results—giving users structured information to assess a model before using it. When a creator adds the not-for-all-audiences tag, Hugging Face automatically displays a warning on that model's page. Users can filter models by license, which is drawn directly from the metadata the creator is required to provide.