Annex IV §2(d)

Annex IV §2(d) — Data Documentation (EU AI Act)

How to document training, validation and test data under Annex IV §2(d) of Regulation (EU) 2024/1689. Data sheets, lineage, special-category derogation, and bias examination evidence.

Source: Regulation (EU) 2024/1689 on EUR-Lex · Last published 2026-04-28 · Draft pending human review

What §2(d) actually requires

Annex IV §2(d) of Regulation (EU) 2024/1689 requires the technical file to document:

"Where applicable, the data requirements in terms of data sheets describing the training methodologies and techniques and the training data sets used, including a general description of these data sets, information about their provenance, scope and main characteristics; how the data was obtained and selected; labelling procedures (e.g., for supervised learning), data cleaning methodologies (e.g., outliers detection)."

This is the technical-file rendering of Article 10. It is where the data sheets, the lineage diagrams, the bias examination, and (where applicable) the Article 10(5) strict-necessity dossier for special-category processing all live.

What to include

A defensible §2(d) entry contains:

Data sheets — one per training, validation and test set. Source, collection process, license/legal basis, demographic distribution, known biases, error rates, completeness.
Lineage diagram — source → ingestion → cleaning → labelling → split → training. Version-control identifiers per stage.
Labelling procedure — annotator profile, inter-annotator agreement, gold-standard sampling, label-quality monitoring.
Cleaning methodology — outlier detection rules, missing-value handling, deduplication, leakage checks.
Bias examination report — per protected attribute (where lawful to identify), per intended-use sub-group: representativeness, statistical-parity / equal-opportunity / calibration-by-group metrics. See Article 10(2)(f)–(g).
Bias mitigation log — for every bias identified, mitigation tried (re-balancing, re-weighting, threshold adjustment, post-processing, architecture change), with before/after metrics.
Article 10(5) dossier if special-category data is processed — strict-necessity argument, safeguards, deletion timeline.
Article 10(4) context analysis — how the deployment context differs from the training context.

A worked structure

Most teams structure §2(d) as:

Subsection 1 — Datasets. A table of train/val/test sets with metadata.
Subsection 2 — Lineage. A diagram + table of stages with code references.
Subsection 3 — Quality controls. Cleaning rules + labelling SOP.
Subsection 4 — Bias examination. Tables + commentary.
Subsection 5 — Mitigations. Before/after metric table.
Subsection 6 — Special-category dossier (if applicable).
Subsection 7 — Context analysis under Article 10(4).

Total length: 5–15 pages, plus data sheets as appendices.

Inline crosswalk

ISO/IEC 42001:2023 Annex A.7.4 — Quality of data for AI systems.
ISO/IEC 42001:2023 Annex A.7.5 — Data acquisition and preparation.
ISO/IEC 42001:2023 Annex A.7.6 — Data provenance.
NIST AI RMF MEASURE 2.2 — Representativeness of data is examined.
NIST AI RMF MEASURE 2.11 — Fairness and bias evaluated and documented.
NIST AI RMF MAP 2.3 — Scientific integrity and TEVV considerations.

Common mistakes

No data sheets. Single most common Annex IV §2(d) failure.
No bias examination by sub-group.
Article 10(5) used as blanket permission instead of strict-necessity carve-out.
Missing Article 10(4) context analysis when deploying outside training context.
No lineage; un-reproducible under Article 74 inspection.

Disclaimer. Reference; not legal advice. Verify with counsel. Reg text from Regulation (EU) 2024/1689.

Reference checklist

From the Governancer 30-item EU AI Act checklist. Each item joins to the ISO 42001 + NIST AI RMF crosswalk table below.

Article 11 · Starter tier · critical

Document training data sources and quality controls

Article 10 data governance must be traceable in the technical file. Where did the data come from, who vetted it, what validation ran.
Article 10 · Pro tier · medium

Document data quality criteria, error rates, and representativeness

Article 10(3) requires datasets be relevant, representative, free of errors and complete. Quantify error rates and coverage gaps.
Article 10 · Pro tier · high

Run bias detection per protected attribute

Article 10(2)(f)-(g) requires examination of possible biases affecting health, safety, and fundamental rights across protected groups.
Article 10 · Pro tier · medium

Maintain end-to-end data lineage from source to training

Provenance chain: source → collection → preprocessing → split → training. Required for reproducibility and audit defence.

ISO 42001 + NIST AI RMF crosswalk

Pulled live from the Governancer crosswalk module. Mapping reference; not a substitute for ISO 42001 certification audit or NIST AI RMF self-attestation.

ISO/IEC 42001:2023

Checklist item	ISO 42001 control	Rationale
`art11-training-data`	ISO/IEC 42001:2023 Annex A.7.4 — Quality of data for AI systems	Article 10 training-data documentation directly evidences the data-quality control objective in Annex A.7.
`art10-data-quality`	ISO/IEC 42001:2023 Annex A.7.4 — Quality of data for AI systems	Article 10(3) error-rate, completeness and representativeness criteria are exactly the data-quality control in Annex A.7.4.
`art10-bias-detection`	ISO/IEC 42001:2023 Annex A.7.5 — Data acquisition and preparation	Bias detection per protected attribute is the substantive control evidenced by data acquisition and preparation in Annex A.7.5.
`art10-data-lineage`	ISO/IEC 42001:2023 Annex A.7.6 — Data provenance	End-to-end source-to-training data lineage is the data-provenance control in Annex A.7.6.

NIST AI RMF 1.0

Checklist item	NIST AI RMF subcategory	Rationale
`art11-training-data`	NIST AI RMF MAP 2.3 — Scientific integrity and TEVV considerations are identified and documented, including data and modeling approach	Documenting training-data sources, quality controls and TEVV is the substance of MAP 2.3.
`art10-data-quality`	NIST AI RMF MEASURE 2.2 — Evaluations involving human subjects meet applicable requirements; representativeness of data is examined	Article 10(3) representativeness, error-rate and completeness measurement is the data-evaluation control MEASURE 2.2 expects.
`art10-bias-detection`	NIST AI RMF MEASURE 2.11 — Fairness and bias — as identified in the MAP function — are evaluated and results are documented	Bias detection per protected attribute, with documented results, is exactly MEASURE 2.11.
`art10-data-lineage`	NIST AI RMF MAP 2.3 — Scientific integrity and TEVV considerations are identified and documented, including data and modeling approach	Source-to-training data lineage is the documented data approach required for scientific integrity under MAP 2.3.

Pro feature

Generate Article 11 with AI

LLM-assisted draft of all eight Annex IV sections, pre-filled from your system intake. 5 drafts/month on Pro.

Pro template

Download FRIA template

15-page Article 27 FRIA template (.docx) with the six elements pre-structured and a worked example.

Get the 30-item EU AI Act compliance checklist

Free PDF. No spam. Maps every Article and Annex IV section we ship to a ready-to-action checklist row.

Reference; not legal advice. Verify with qualified counsel before relying on it for compliance decisions. Reg text quoted from the Official Journal version of Regulation (EU) 2024/1689. Published by Agonist Development AB.