Resources
Dataset manifest vs datasheet vs dataset card
A manifest identifies files. A datasheet records structured answers about a collection. A dataset card explains the collection to users. Most serious deliveries benefit from all three.
Published 2026-10-28 · 5 min read
A manifest identifies files. A datasheet records structured answers about a collection. A dataset card explains the collection to users. Most serious deliveries benefit from all three.
Use the linked free tool to work through the practical steps. Keep the original files and documentation unchanged, and save your report with the date and scope of the review.
The manifest is the file-level record
Give each delivered file a stable relative path or identifier. Include file size, measured duration where relevant, format and the metadata required by the project. For recording collections, session and view identifiers often belong beside technical properties.
The Media Dataset Inspector produces a technical file manifest. It does not generate transcript labels, session relationships or participant consent records. Add those through a documented mapping process rather than guessing them from filenames. If checksums are required, calculate and validate them separately; this tool does not compute hashes.
The datasheet is a structured collection description
A datasheet can be a question-and-answer table covering source, purpose, collection method, annotations, quality checks, permissions and limitations. It helps reviewers compare collections using consistent questions.
Fiund’s builder exports those answers as CSV. That is a collection datasheet, not a per-file manifest. A field being filled out only means an answer was supplied; it does not mean the answer has been verified. Keep evidence references and reviewer decisions in your diligence process.
The card is the reader-facing explanation
A dataset card typically presents the collection in a readable Markdown document. It can include the same information as a datasheet while giving users context on appropriate uses and limitations. Hugging Face uses README files as dataset cards and recognizes YAML metadata for discovery.
Keep the public card useful without exposing private evidence. For commercial collections, describe how an eligible buyer can request evaluation and documentation. Do not present a public card as permission to download or use files beyond the applicable terms.
Keep the three documents consistent
Use one dataset identifier and version across the card, datasheet and manifest. Reconcile counts and units before delivery. If a card says ten unique activity-hours while the manifest totals forty camera-file hours, explain the four-view setup rather than silently substituting one figure for the other.
After updates, regenerate the inventory and review the card. Track removals, replacements and changed annotations. A consistent package helps reviewers investigate exceptions and makes future delivery checks repeatable. Start with the free builder for documentation and the media inspector for file-level properties.