Sensitive Span Detector (experimental)
An experimental model that flags candidate sensitive spans in text for downstream redaction. It supports, but does not replace, human review.
Intended use
- Assisting a privacy-preserving pipeline by proposing spans for review.
- Research into detection/redaction trade-offs.
Out-of-scope use
- Sole, unreviewed redaction of real personal or health data.
- Any use implying a guarantee that all sensitive data is removed.
- Clinical or legal decision-making.
Metrics
No verified metrics are published for this model.
Model type
Token-classification model for sensitive-span detection.
Architecture
A fine-tuned transformer encoder. Details are provisional and may change.
Training data
Synthetic and public-source text with sensitive-span annotations. No real personal or health records were used.
Performance note
Quantitative performance is not published because it has not been validated on a representative, licensed evaluation set. Reported numbers would be misleading at this stage.
Ethical considerations
Presenting this as a complete privacy solution would be harmful. It is a research aid that must sit behind human review and residual-risk evaluation.
Privacy considerations
The model itself must not be trained on, or used to memorise, real sensitive data.
Limitations
- Recall is not guaranteed; sensitive spans will be missed.
- Trained largely on synthetic data; real-world distribution shift is expected.
Known failure modes
- Under-detection of rare or novel sensitive formats.
- Over-detection causing unnecessary redaction of benign text.