Developer fine-tunes 1.7B AI model for medical data redaction using only synthetic data

A developer has fine-tuned a 1.7-billion-parameter language model to perform PHI (Protected Health Information) de-identification without using any real patient data at any stage of the process. The model serves as the second-stage adjudicator in 'localscrub', a local-first privacy tool that handles ambiguous spans flagged by a rules engine. Training data was generated entirely through a synthetic corpus tool that plants identifiers into fake clinical notes, ensuring perfect labels by construction and zero privacy exposure. Two thousand training examples were produced in seconds using a single command, with separate random seeds used for training and evaluation sets to prevent data overlap. The developer acknowledges that because only seven note-skeleton templates underpin the corpus, results on template-based benchmarks are optimistic, and performance on authentic medical transcription data serves as the more credible generalization test.