A housing authority wants to fine-tune a foundation model on historical casework notes to help staff draft case summaries, but the notes contain resident names and home addresses. What should happen to this text before it is used as fine-tuning data?

Mechanism: masking or anonymizing personally identifying fields like names and addresses replaces or removes sensitive values while preserving the surrounding narrative content the model needs to learn from, this matters because a fine-tuned model can memorize and later reproduce verbatim training examples, including PII, when prompted a certain way. Why the wrong options fail by concept: assuming a training job's isolation prevents exposure ignores that the risk isn't the job environment, it's what the resulting model learned and can later output, isolating the job doesn't isolate the model's memorized knowledge. Categorical encoding transforms discrete labels into numeric representations for structured features, it isn't a text-redaction technique, and applying it to free-form casework narrative would mangle the content the model actually needs to learn from. Deduplicating notes reduces redundant records, but it does nothing to remove identifying details from the records that remain, a resident's name in a unique note is still exposed after deduplication. Scope caveat: anonymization needs to be applied consistently across the whole dataset, including any cross-references between notes, since combining partially redacted fields across records can still allow re-identification. Operational check: run a sample of the fine-tuning corpus through a PII-detection pass after masking to confirm no names or addresses remain before the data enters the training job.