liveBook · Manning

PII scrubbing

As explained in LLM Customization and Fine-Tuning

The removal or masking of personally identifiable information from data.

PII scrubbing removes or masks personal information before data is reviewed, annotated, stored for training, or used to train a model. Training on personal identifiers can lead a model to reproduce them and creates data-processing liability. section 3.8

Scrub both direct identifiers and combinations that could identify someone:

  • Structured identifiers include names, emails, phone numbers, national identifiers, account numbers, IP addresses, and device identifiers. Automated tools can detect these and replace them with typed placeholders. section 3.8
  • Indirect identifiers can emerge from combinations such as age, job title, and employer. Automated detection may miss these; human review of a sample is needed. section 3.8

Scrubbing production logs

Before reviewers inspect exported query-and-response pairs, scrub the data. Microsoft Presidio can detect more than 20 entity types and replace them with placeholders such as [PERSON] and [EMAIL]. Then review at least 10% of examples for indirect identifiers and document the review. section 3.4.4 section 3.8

How LLM Customization and Fine-Tuning covers it

How deep it goes
Distinguishes structured from indirect identifiers and recommends automated scrubbing followed by documented human review of at least 10% of examples.
Examples
Uses Presidio placeholders such as [PERSON] and [EMAIL], and identifies combinations like age, job title, and employer as potential indirect identifiers.
What you can do after
Apply Microsoft Presidio for structured identifiers and document human review of a sample for indirect identifiers.
How it connects
Connects data governance, production log mining, model training, and human review of indirect identifiers.

Also explained in

Compare how 2 books explain PII scrubbing →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Customization and Fine-Tuning. The section links point to the parts of the book it draws on.