Training Dataset Construction and Release

A training dataset should begin with a spine: one row for every example the learning protocol intends to include. The spine carries the example key, the subject, the prediction time, and often the outcome window. Features are then joined onto that spine under explicit temporal and cardinality rules. Labels are assigned separately, so that feature availability and outcome logic both stay inspectable. Starting from a large feature table can reverse this discipline. The available rows begin defining the population. Examples with missing joins disappear without a deliberate exclusion rule.

A published training dataset is therefore a release artifact. It combines a population definition, source snapshots, transformation code, feature availability, labels, splits, diagnostics and documentation. The rest of this lesson takes those parts one at a time. For each it asks the same two questions: what does a published dataset that actually does this look like, and what does the record show when it is skipped?

Separating stages makes population changes and target logic easier to inspect.

Stage 3 is the one most often treated as a lookup rather than as a product with its own error rate. The test sets of 10 widely used vision, language and audio benchmarks are the sets whose numbers the field reports as ground truth. Northcutt and colleagues tested them for label errors, and opened their abstract with the result: “Errors in test sets are numerous and widespread: we estimate an average of at least 3.3% errors across the 10 datasets, where for example label errors comprise at least 6% of the ImageNet validation set.” The algorithmically flagged candidates were not taken on trust either. Crowdworkers confirmed 51% of them as erroneous.

The figure has since been repeated from outside the group. Anzaku and colleagues, re-annotating ImageNetV2, write that “Label errors have been identified within the test sets of numerous commonly used datasets, including a 6% error rate in the ImageNet-1k validation set [11].”

The most scrutinised label sets in the field carry that much error. So a label column assembled in-house, out of an operational review process, is not a fact to be joined in silently at stage 3. It is a stage that owes the reader a count.

The right choice depends on data scale, reproducibility needs, and how often source history changes.

Publishing a version should make its differences from the last one reviewable. It should also make the terms of use, the known defects and the exit route explicit.

A standard form for this was proposed in 2018. The datasheet, from Gebru and colleagues, asks every dataset to document “its motivation, composition, collection process, recommended uses, and so on”. The analogy is the electronics datasheet that ships with each component. The proposal reached Communications of the ACM in December 2021.

Since 1 August 2024, when it entered into force, that documentation has had a statutory floor. Regulation (EU) 2024/1689, the AI Act, states at Article 10(3): “Training, validation and testing data sets shall be relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose.” Article 10(2)(f)–(g) add the examination for possible biases and the measures to detect, prevent and mitigate the biases so identified. The European Commission's own regulatory framework page puts the high-risk requirement as “high-quality of the datasets feeding the system to minimise risks of discriminatory outcomes”. For a high-risk system in the EU, step 5's intended use, restrictions and known gaps are not house style. They are obligations under a named regulation, with an article number.

Step 5's deprecation policy is not hypothetical either. On 23 December 2023 David Thiel of the Stanford Internet Observatory published an evaluation of LAION entries above a safety cutoff, 32,138,129 items examined. His report says what came back: “Through this process, we identified 3,226 dataset entries of suspected CSAM, much of which was confirmed as CSAM by third parties.” LAION withdrew the corpus. On 30 August 2024 it republished as Re-LAION-5B, and the release note gives a removal count rather than a reassurance: 2,236 links removed after matching against hash lists from the Internet Watch Foundation (18), the Canadian Centre for Child Protection (1,129) and Stanford (1,714). That set subsumes the 1,008 links named in the December 2023 report. A named version, a withdrawal, a successor with a new identifier, and an itemised list of what left: that is step 5 being exercised rather than promised.

Compare dataset versions at several levels. Start with example keys: added, removed, and retained rows. Then compare labels, features, missingness and important slices among the retained examples. Separate expected changes from unexplained ones. A new source may add coverage intentionally, while a sudden category collapse may signal a producer bug. Model metrics alone are too downstream for this task. A model can hide a data defect temporarily, and fixing a real data defect can lower an overoptimistic score.

MIMIC-IV publishes this comparison in public. It is distributed through PhysioNet by the MIT Laboratory for Computational Physiology, and it has appeared as eight dated versions: 0.3 and 0.4 on 13 August 2020, 1.0 on 16 March 2021, 2.0 on 12 June 2022, 2.1 on 16 November 2022, 2.2 on 6 January 2023, 3.0 on 23 July 2024 and 3.1 on 11 October 2024. The v3.0 release notes state the population delta in figures rather than in adjectives: patients 299,712 → 364,627, admissions 431,231 → 546,028, ICU stays 73,181 → 94,458. The reason is one sentence: “Stays occurring between 2020 and 2022, inclusive, were added to the database.”

The pre-v3.0 side of that diff is independently checkable. Johnson and colleagues, in Scientific Data, describe 431,231 hospital admissions and 73,181 ICU admissions collected between 2008 and 2019. A team whose metrics moved after upgrading can therefore ask a specific question — did the added years change the case mix? — instead of a vague one.

Store the diff report with the release. It becomes the first document investigators consult when training behavior changes after a data update.

Each step needs an explicit policy and a diagnostic count. The cost of leaving them implicit has been counted across the sciences. Kapoor and Narayanan surveyed reviews of ML-based science and summarised the result this way: “Through a survey of literature in fields that have adopted ML methods, we find 17 fields where leakage has been found, collectively affecting 294 papers and, in some cases, leading to wildly overoptimistic conclusions.” Their paper appeared in Patterns on 4 August 2023, and it sorts the failures into a taxonomy of eight distinct leakage types. The earlier preprint, from 14 July 2022, had put the count at 329 papers across the same 17 fields.

The finding has been restated from outside that group. Simon Roth, surveying 2,047 benchmark datasets, writes: “Kapoor and Narayanan audited the machine learning literature across 17 scientific fields and found 294 published papers whose results were invalidated by data leakage after publication; their living survey now catalogues 648 papers (as of mid-2024) across 30 fields.”

None of those 294 papers set out to define a population by accident. The five steps below are how it happens.