Source: Hugging Face Hub docs: Model Cards
How Hub model cards work: README.md with YAML metadata for license, base_model and relation, datasets, pipeline_tag, library_name and eval results.
Contents
- What: official documentation
- Publisher: Hugging Face
- Date: accessed 2026-11-14
- Original: Model Cards
Summary
A model card is the README.md of a model repo. It combines prose (intended use, limitations, training data and setup, evaluation) with a YAML metadata header. The Hub uses that header for search filters, license display, lineage and widgets.
Key claims
- A card should describe the model, its intended uses and limitations (including bias and ethical considerations, following Mitchell et al. 2018), training parameters, datasets and evaluation results.
- The
licensefield takes a standard identifier. For a custom license, setlicense: otherand addlicense_nameandlicense_link. base_modelrecords lineage, as one model ID or a list for merges. The Hub infersbase_model_relation(adapter, merge, quantized, finetune), and authors can set it explicitly.- Repos created after August 2024 must set
library_name: transformersexplicitly. Aconfig.jsonfile no longer implies it. pipeline_tagsets the task, which drives filtering and the inference widget.datasetslinks to training data.new_versionpoints to a successor model.- Evaluation results can be machine-readable in
model-index, a format derived from Papers with Code, or in a newer, simpler eval-results format. - Linking an arXiv paper automatically adds an
arxiv:<id>tag.not-for-all-audiencesputs a warning interstitial in front of the card.