Preparation turns raw data into a usable dataset: exploring, cleaning, filling gaps, scaling, composing, splitting, labelling and de-identifying.
What the base framework establishes
The clause says data is rarely ready for training and that preparation can take a large share of the effort. It lists the activities: exploring the raw data to understand its shape and distribution; wrangling it into a structured table; cleaning duplicates, missing values, invalid entries and outliers; imputing missing values; normalising or scaling numeric features; composing a dataset from several sources; splitting it into training, validation and test sets that do not overlap; labelling samples; and de-identifying personal data. It also names data poisoning as a check, because contaminated training data can cause harmful behaviour, and it notes that incomplete or unrepresentative data can produce biased or discriminatory predictions.
Why it matters to an AI management system
Preparation is where most silent defects enter. A split that leaks, a label rule applied inconsistently or an outlier removed without a note all change the result and leave no trace unless the record captures them. The clause also puts de-identification and poisoning in the same stage as cleaning, which is the right place for both.
The public draft summary names no change to 8.3 Data preparation. The final text could still touch it.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Preparation log | Each step applied, in order, with the parameter and the reason. |
| Split and leakage check | The split method and the test for overlap between sets. |
| De-identification record | The method, the fields affected and the residual risk. |
One question to prepare for. Could you reproduce this dataset from the log, and does the test set leak?
Common gaps
- The preparation notebook is the only record and it is not versioned.
- Rows are split randomly after duplicate removal, but near-duplicates remain across sets.
- De-identification is asserted with no method and no residual risk note.
Prepare before publication
Turn the preparation notebook into a dated log with parameters and reasons. Add a near-duplicate leakage check and a de-identification record.
Dates that matter
No separate date attaches to 8.3; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
3.3 Data · 8.2 Data acquisition · 8.4 Modelling · 8.5 Verification and validation
Source
Source for 8.3: ISO/IEC 23053:2022, clause 8.3. The public sample stops before 8.3, so this page relies on the public contents and the public summary and does not quote the clause. The two official records are linked on this page so any line can be checked.
Not legal advice; does not certify. A reading of the framework and the evidence a reviewer commonly samples, not a substitute for the standard or your auditor's rules.