liveBook · Manning

Personal data

As explained in Building Specialized AI Systems

Personally identifiable information is information that can identify a particular individual, directly or through combination with other data.

Personal data is information that can identify an individual directly or when combined with other data. In fine-tuning workflows, it is treated as a privacy and compliance risk requiring removal or masking before use. section 2.4.1

Handle personal data during dataset preparation by:

  • Identifying sensitive fields such as client names, account identifiers, and client-level transaction records. section 2.4.1
  • Removing or masking identifying values through de-identification before fine-tuning. section 2.4.1
  • Using synthetic data when testing pipelines without exposing real individuals. section 2.4.1
  • Checking for remaining sensitive information as part of a broader data-readiness checklist. section 2.4
Listing 2.19 Cleaning example
import re
import pandas as pd
 
data = { #A
    "account_id": ["123-456-789", "987-654-321"],
    "client_name": ["Alice Johnson", "Bob Smith"],
    "transaction": ["USD 1,000", "$2,500.00"],
    "notes": [
        "Purchase at Grocery Store. CONFIDENTIAL.",
        "Wire Transfer. DISCLAIMER: For internal use only."
    ]
}
 
df = pd.DataFrame(data)
 
df["account_id"] = df["account_id"].apply(lambda x: "[ACCOUNT_ID]") #B
df["client_name"] = df["client_name"].apply(lambda x: "[CLIENT]") #B
 
def normalize_amount(value): #C
    cleaned = re.sub(r"[^\d.]", "", value)
    return float(cleaned)
 
df["transaction"] = df["transaction"].apply(normalize_amount)
 
df["notes"] = df["notes"].apply( #D
    lambda x: re.sub(r"DISCLAIMER:.*|CONFIDENTIAL\.", "", x).strip()
)

Illustrates masking account identifiers and client names while removing confidentiality and disclaimer artifacts from transaction notes. section 2.4.1

How Building Specialized AI Systems covers it

How deep it goes
Explains sensitive-data risks, masking, synthetic substitutes, and checklist placement without detailing privacy law or re-identification testing.
Examples
Uses client names, account identifiers, client transactions, and masked financial transaction records as examples.
What it assumes
Assumes familiarity with datasets and fine-tuning while introducing privacy handling through finance-oriented examples.
What you can do after
Apply de-identification and synthetic-data techniques when preparing privacy-sensitive fine-tuning datasets.
How it connects
Connects personal data with de-identification, synthetic data, data cleaning, dataset curation, and fine-tuning.

Also explained in

Compare how 13 books explain personal data →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Building Specialized AI Systems. The section links point to the parts of the book it draws on.