Data Set

A dataset is a collection of data organised and distributed as a unit. The definition is uncontroversial until something has to be done with it, at which point the interesting question turns out to be where its boundary is drawn.

That question is not academic. A dataset is the unit that gets a DOI, a licence, a citation, and a metadata record. Draw the boundary too wide and the citation is uninformative — “the model output” says nothing about which part was used. Draw it too narrow and every analysis cites hundreds of objects and nobody maintains them.

Where the boundary actually goes

For observational data the natural unit is usually a station or network over a period, and it holds up reasonably well.

For model output it is genuinely unclear. Is the dataset one simulation? One ensemble member? All variables at all levels, or one variable? A single WRF downscaling experiment can produce thousands of files across variables, levels, and time slices, and any grouping of them is a defensible dataset. The convention that has emerged is roughly “what a user would want to cite as a unit”, which is circular but works in practice better than more principled alternatives.

The practical rule I have settled on: the dataset is the smallest unit that is independently useful. A single variable without its coordinate information is not, so it is not a dataset. One variable across the full run usually is.

What makes one reusable

Beyond the data, a dataset needs enough context that someone else can interpret it correctly rather than plausibly:

• Variable definitions and units, unambiguously — this is what CF conventions exist to standardise.
• Spatial and temporal coverage and resolution, including the spatial reference system, which is skipped surprisingly often.
• Provenance — what produced it, from what inputs, with what version.
• Uncertainty or quality information, which is the most commonly missing item and frequently the one that determines whether the data can answer the question at all.
• A licence. Data with no licence is legally unusable, whatever the intention behind releasing it.

Data without these is not a dataset so much as an archive of numbers. It can be stored, and it can be pointed at, but it cannot be reused by anyone who was not there when it was made — which is the situation FAIR principles exist to address.

See also: climate data for the domain-specific case, metadata, FAIR, and DOIs for citation.