Zovda
Most data cards describe the good news: volume, languages, a list of categories. The questions that decide a purchase come after those numbers are accepted.
A card that answers only the good news is a brochure with a table in it.
The six questions
- What unit is one row, exactly
- Which fields are always present and which are sometimes empty
- What is deliberately not in the set
- How was it filtered, by which rules, in which order
- How were disagreements between labelers resolved
- Which slice should not be used for evaluation, and why
The third and sixth are the ones buyers ask about and sellers omit.
Numbers need a denominator
"Two thousand hours" means nothing without saying whether that is before or after the segments that failed quality review. "Nine languages" means nothing without the share of each.
Put the denominator in the same cell as the number. It prevents the sentence that starts with "so that is about" in a meeting two weeks later.
Known gaps belong on the card
Every collection has a hole. Naming it builds more trust than any claim of completeness, because the buyer has already found it and is now wondering what else was left out.
A gap written down is a scoping conversation. A gap discovered later is a support ticket.
Keep the card next to the manifest
The card should be generated from the same source as the field list, or the two will drift and the buyer will trust neither. When the schema changes, the card changes in the same commit.
Zovda ships each collection through Datuva with a card that records how it was cut and what it excludes. The image collection shows the field list a card is built from.
Tell us what you need to verify before you buy. We reply within one business day.