clinanon-student

Token classifier that finds privacy-related information in clinical dialogues and notes: names, places, organizations and occupations (it was also trained on ages, dates, times, contact details and IDs, but clinanon uses the rules for those). It is the statistical detector of clinanon, built for the Bosch challenge at Neural Wave 2026, and is meant to be used inside that pipeline, behind its clinical veto. It is not a standalone de-identifier.

• Base model: StanfordAIMI/stanford-deidentifier-base (BERT-base, uncased, MIT), runs on CPU.
• Training data (distillation): Qwen3.8-27B (Apache-2.0) acted as teacher.
  • It labelled 140 ACI-Bench encounters (CC BY 4.0, with ICD-10 annotations under CC BY-NC 4.0).
  • It wrote and labelled 1,082 synthetic encounters, all fictitious.
  • The labels were filtered with clinanon's clinical veto.
• Labels: BIO tags for NAME, AGE, DATE, TIME, LOCATION, ORGANIZATION, PROFESSION, CONTACT and IDENTIFIER.
• Results:
  • span F1 0.974 against the teacher on held-back documents
  • inside clinanon, 94.4 % of private details removed on a synthetic test set (90.4 % without the student)

Licensed CC BY-NC 4.0 because part of the training data is.