Before you adopt a model, read its card for the license, datasets, base model, intended uses and limits, and evaluation results. On the Hugging Face Hub a model card is the repository README with a YAML metadata block at the top. Treat blanks as open questions to resolve, not as a clean bill of health.
What a model card is
The Hub documentation describes model cards as files that accompany models and provide information. They are Markdown with extra metadata, rendered from README.md. The card should describe the model, its intended uses and potential limitations including biases and ethical considerations, the training parameters, the datasets used, and evaluation results.
The metadata also drives Hub features such as filtering and showing the license.
Fields to read, in order
license. The metadata can hold a license identifier. If the author used the value other, a name and link may follow. Read the actual text. If the license is missing, find out before using the model commercially.
datasets. Authors can list datasets used for training, and the Hub links them. An empty list may mean the author did not say, not that no data was used.
base_model. Shows whether this model is a fine-tune, an adapter, a quantized version or a merge of others. Each parent brings its own license and risks. See the lineage post linked below.
library_name and tags. Tell you which library is expected to load it, which tells you what code path will run.
Intended use and limitations. Read for what the author says the model should not be used for.
Evaluation results. The Hub can parse and display structured results. Treat them as the author's claims. Run your own tests on your own data.
new_version. The card can point to a newer version. Useful when you want to know whether the one you adopted is still maintained.
What the card cannot tell you
A card is written by the author and not verified by the Hub. It does not prove the files match the description, and does not replace scanning, signing or testing. A careless or malicious author can write anything. Cross check what matters, such as license and lineage.
A five minute routine
- Open the card, note the date and the revision you read.
- Copy license, base model and datasets into your intake record.
- Write down every blank.
- Check the files tab for the weight format.
- Decide: approve, ask the author a question, or reject.
Related: base model metadata explained.
Write down what you could not find
Most real cards are incomplete. Some skip the datasets, some skip limits, some have a license that only says other. Copy those gaps into your intake record so that a later reviewer sees the question was asked. If the model is important to your product, send the author a short email or open a discussion on the repository asking for the missing facts, and note the date you asked.
Quick check
Is a model card verified by the Hub?
The docs describe it as a README with metadata that the author writes. Treat it as a claim to cross check, not a certification.
What if the license field is empty?
Treat the license as unknown. Ask the author, or avoid commercial use until you have written confirmation.
Sources
Last reviewed: 19 October 2026.
