We want to fine-tune the generator on 200,000 support t...

A regex-based approach is insufficient because it only removes formatted subsets like email addresses, leaving unformatted personal data—such as names, employers, and cities—intact. Because duplication is what makes data extractable, a customer appearing in multiple tickets is at high risk. The better approach is to use synthetic transcripts for tone transfer, keeping the real data in a datastore where it can be managed and deleted via index operations.