Personal data is information relating to an identified or identifiable individual. In model fine-tuning, it creates a training-data risk: a model can learn identifiers from examples and reproduce them at inference time. Removing the identifiers from the dataset after training does not remove them from the model. section 3.8 section 3.9
The default governance decision is to identify and remove personal data before training examples reach annotators, annotation platforms, vendors, or training infrastructure. Retention requires a task-specific justification and must be permitted by policy and law. Privacy and data-residency requirements can also constrain where fine-tuning occurs. section 3.8 section 3.9 section 5.11
Review data at two levels:
- Structured identifiers include names, email addresses, phone numbers, national identifiers, financial account numbers, IP addresses, and device identifiers. Microsoft Presidio combines regular expressions and named entity recognition to detect entity types and replace them with placeholders such as [PERSON] and [EMAIL]. section 3.8
- Indirect identifiers are combinations of attributes that can identify someone even when no single attribute does so. Automated scrubbing needs human review for these combinations. section 3.8
An identifiable combination
Age, job title, and employer may each appear innocuous, but together they can identify one person in a small organization. A review must consider the combination, not just whether each field matches an identifier pattern. section 3.8
A practical workflow is to scrub automatically, then review at least 10% of examples by hand and document the review for an audit trail. For healthcare data, HIPAA Safe Harbor specifies 18 identifier types for removal and calls for assessing whether residual data could identify an individual. section 3.8












