Skip to main content

Datasets: Provenance, Leakage, and Evaluation

A downloadable table is not automatically evidence. Before the 1936 US presidential election, The Literary Digest polled about ten million people, received about 2.38 million replies, and predicted a landslide for the Republican Alf Landon; Roosevelt then carried every state except Maine and Vermont. The list came from the magazine's own readers, car registrations, and telephone subscribers. Squire's 1988 study found that the sample was biased and that who chose to reply was biased too, and that together they produced the badly wrong estimate. Two million replies looked like a lot, but they stood only for people who were on those lists and chose to answer.

A dataset therefore records one collection regime, measurement process, version, and scope. Both missingness and inclusion carry the bias of the collection mechanism: the Digest included listed people who replied and missed every other voter.

Fitness Review​

DimensionAskCommon failure
provenancewho produced it, why, and how?a mirror loses definitions and version
unitdoes a row mean person, event, window, or aggregate?repeated entities treated as independent
populationwhom should conclusions cover?convenience sample called representative
measurementhow are variables and labels observed and revised?an administrative proxy treated as the thing to be measured, such as health costs standing in for how sick patients are: equally sick Black patients had less spent on them, so the algorithm judged them healthier (Obermeyer et al., 2019)
timewhen are features and labels available?future fields or revised values leak
missingnesswho is absent and why?imputation hides systematic invisibility
joinsare keys unique and at one granularity?many-to-many join multiplies rows
rightsdo license, consent, and privacy permit use?public visibility treated as permission
reproducibilitycan version, query, checksum, and transforms be rebuilt?URL points only to changing latest data

Rejecting an unsuitable dataset early is a successful result.

Once the data fits, fix the split with Data Splits and Leakage before any processing learns from the data; that page also ends by separating data quality, leakage, and distribution shift. To tie test results to the people and costs of actual use, see Evaluation Under Distribution Shift.

Minimal Datasheet​

Datasheets for Datasets gives the full framework. For an experiment, record five items first: the name and an immutable version; the source, retrieval date, and query or checksum; the ordered, executable transformation steps; the split keys (entity, group, or time); and the supported and unsupported uses. Give each remaining row of the table one sentence. With these, an experimental result can be traced back to an exact data version and its processing steps.

The Dataset You Build Most Often: An Evaluation Set​

The dataset an individual developer most often builds by hand is an evaluation set drawn from their own prompts and logs: a few dozen to a hundred real requests, each with an acceptance criterion that can be checked from outside. The table above still applies. Decide first whether a row is one request or a whole conversation; several similar requests from one user count as one entity and go on the same side of the split.

Check the rights and privacy of the logs separately. Logs often hold customers' names, email addresses, contract details, or private text someone sent you. Putting that text into an evaluation set, or sending it to a third-party evaluation API or grading model, is a different use, and the original terms and the other person's consent have to cover it. After sampling, remove or replace whatever identifies a person before archiving; when the original text must be kept, keep it on your own machine and out of public repositories. For how to run and record the evaluation, see evaluating a model on your own tasks.

Stop Conditions​

Do not train a complex model while prediction time, target, entity boundaries, label maturity, rights, split key, baseline, or error costs remain undefined. Record unknowns as blockers rather than hiding them behind a default random split. Even a complete datasheet only makes the experiment traceable; it does not certify ethics or statistical validity.

Explore connectionsOpen network