JACA369/sensitivity-scanner-v6

Sensitivity Scanner v6 (browser build)

A small model (74M parameters, 143 MB fp16 ONNX) that flags English text containing sensitive personal information. Its purpose is classifying sensitive university files, such as documents in a university Google Drive, so that student records, medical details, ID numbers, financial details, credentials and HR matters can be audited. It's a fine-tuned GLiNER2.5-small (fastino/gliner2.5-small-v1), exported to ONNX so it runs entirely in the browser with ONNX Runtime Web.

It classifies a block of up to 100 words into one of 7 labels:

Label	Meaning
medical	Health, medical, or mental-health details about a specific person
student_records	Grades, conduct, advising, accommodations, or other FERPA records about a named student
id_numbers	Social Security, passport, driver's license, visa, or other ID numbers
financial	Bank accounts, card numbers, salaries, income, or payroll for a specific person
credentials	Real passwords, API keys, private keys, or login secrets
legal_hr	Legal, HR, disciplinary, harassment, criminal, or immigration matters about a person
none	None of these: general, public, aggregate, fictional, redacted, or placeholder content

A block's score is 1 − P(none), and a file is flagged at the first block scoring 0.5 or higher. The model is meant to be used with pattern rules for structured secrets (SSNs, card numbers, API keys) that run first.

Files

File	What it is
onnx/model_fp16.onnx	The model (fp16). Gives the same decisions as the PyTorch checkpoint on every test block.
tokenizer.json, tokenizer_config.json, config.json	Tokenizer and encoder config
scanner_spec.json	The label prompt as token IDs, label order, threshold and block size. The browser appends a block's tokens after the prompt.

Results

On 647 English test files with the full pipeline (rules + model):

• Missed: 0 of 309 high-sensitivity files
• False alarms: 40 of 317 harmless files
• Speed: about 0.06 s per file on a laptop CPU

Training data

8,983 labeled blocks: synthetic templates (fictional schools and people), plus public datasets: Nemotron-PII, Gretel synthetic PII finance, MTSamples, Asclepius synthetic clinical notes, Wikipedia, AG News and legal-contract sections. No real personal data about real people was used for the sensitive examples.

License

CC BY-NC-SA 4.0, non-commercial use only. The training data includes 300 blocks from Asclepius (starmpcc/Asclepius-Synthetic-Clinical-Notes), which is licensed CC BY-NC-SA 4.0. Check the base model's license (fastino/gliner2.5-small-v1) before any other use.

Limitations

• English only.
• Tuned for recall: about 1 in 8 harmless files is flagged.
• Some test sets were used while writing the pattern rules, so real-world results may be lower than the numbers above.