Housing caseworkers need case notes ready for a fine-tune job. Which data-preparation step best reduces sensitive-data exposure before training starts?

Secure data preparation is a mandatory lifecycle gate before fine-tuning on municipal case content. Housing case notes often contain names, addresses, income details, household composition, and other personally identifiable information. Redacting that PII during preparation—before fine-tuning begins—reduces the chance that the resulting model memorizes or regurgitates unique resident identifiers and shrinks the blast radius if training artifacts, logs, or checkpoints are later exposed. Preparation-time redaction also keeps sensitive fields out of labeling tooling and vendor processing paths that may sit outside the city's primary system of record. Encrypting personally identifiable case notes in storage during preparation protects data at rest, but it does not redact PII from the fine-tune corpus, so resident identifiers remain exposed to memorization once decrypted for training. Restricting personally identifiable case notes with roles during preparation limits who can view the data, but it does not remove PII before training, so sensitive identifiers can still be learned by the model during fine-tuning. Pseudonymizing identifiable case notes with reversible tokens during preparation replaces direct identifiers but leaves a reversible mapping, so residents can still be re-identified and token patterns remain exposed to memorization. Exam caveat: data minimization and de-identification belong in preparation, not as an afterthought once models are distributed. Operational check: run automated PII scanners plus human review on housing note extracts, retain a redaction log for audit, and refuse to start fine-tune jobs until the prepared corpus passes a no-raw-identifiers gate.