Jayant9928/PII-Gate-Nemotron-Anyjev

PII/PHI Gate (Qwen3.5-4B, AnyJev L2) — trained on Nemotron-PII

A calibrated "does this text contain PII/PHI that must be masked?" decision model. It answers in one forward pass with no text generation and returns a probability you can threshold — so it can sit in front of an expensive redaction / de-identification step and skip text that confidently contains nothing to mask.

It is a closed-form linear head on the hidden state of Qwen/Qwen3.5-4B, built with AnyJev. The base model is unchanged; this repo is only the head (~55 KB). Trained only on the open, fully synthetic NVIDIA Nemotron-PII dataset (CC BY 4.0).

License (this head)	CC BY 4.0
Training data	nvidia/Nemotron-PII only — fully synthetic, CC BY 4.0 (© NVIDIA)
Base model	Qwen/Qwen3.5-4B — unchanged, under its own license (download it separately)
Method	AnyJev L2: closed-form linear head on the hidden state, no fine-tuning
Real personal data used	None

Triage, not redaction. "No" means "safe to skip at your chosen risk level", never "this text is anonymised".

Results — Nemotron-PII test split

4,976 held-out examples (train and test records never overlap), 50% containing PII/PHI; positives and negatives are length-matched so text length cannot give the answer away.

Target missed-PII rate	Skip when P(PII) ≤	Texts skipped	Share of the PII-free texts skipped
≤ 0.1%	0.0039	25.4%	51%
≤ 0.5%	0.0297	40.9%	81%
≤ 1%	0.0545	43.8%	87%
≤ 2%	0.1899	48.4%	96%

Accuracy at 0.5: 96.8–97.1% · expected calibration error 0.010 · ~65 decisions/s on one GPU.

Out-of-domain caution. These numbers are in-distribution. On document styles unlike Nemotron-PII — for example template-heavy EHR / XML exports — the head transfers much less well (far fewer texts can be skipped at the same miss rate). Measure on a labelled sample of your data and pick your threshold there before relying on it.

Training data and label mapping

• 13,935 examples from the Nemotron-PII train split: whole records and their natural pieces (paragraphs / form sections), 50% with and 50% without PII/PHI, length-matched, deduplicated; 30 industries, structured and unstructured, US and international.
• A text is positive if it contains at least one span of a label that must be masked. 38 of the 55 Nemotron labels count (names, date of birth, contact details, street address / city / county / postcode / coordinates, and every identifier: medical record, health plan, SSN, national / tax / customer / employee / account / card / routing numbers, licences and plates, device / biometric / vehicle IDs, IP / MAC, usernames, passwords, PINs, API keys, CVV, cookies); age counts only for 90+. Not counted: ordinary dates and times, country, state, company, occupation, education, employment status, gender, race/ethnicity, religion, language, blood type, political view, sexuality. Full mapping: label_mapping.json.
• Head, layer and regularisation chosen by AnyJev's 5-fold cross-validation (out-of-fold accuracy 96.8%).

Usage

import json
from huggingface_hub import hf_hub_download
from anyjev import Decider, Question
from anyjev.backends.hf import HFBackend

art = json.load(open(hf_hub_download("Jayant9928/pii-gate-qwen3.5-4b-nemotron", "contains_phi.head.json")))
Q = Question.noul(art["text"], name="contains_phi")
gate = Decider(HFBackend("Qwen/Qwen3.5-4B"), level="L2")
gate.load_artifact(Q, art)

p = gate.decide_batch(["The change control process is reviewed every quarter."], Q, level="L2")[0].distribution["Yes"]
print("skip" if p <= 0.0297 else "send to redaction")

See example.py. Requires anyjev>=0.2.0 and Qwen/Qwen3.5-4B. Use the exact question text stored in the artifact.

Limitations

• English only; synthetic training data; validate on your own data before choosing a threshold.
• A missed-PII rate is never zero — keep full redaction on everything the gate does not skip.
• One question, one base model: the head works only with Qwen/Qwen3.5-4B and this question text.
• Not a compliance determination (HIPAA, GDPR or otherwise).

Attribution and license

• Training data: NVIDIA Nemotron-PII — https://huggingface.co/datasets/nvidia/Nemotron-PII — licensed under CC BY 4.0. The data was filtered, split into pieces, relabelled to a binary PII/PHI target and length-matched as described above.
• This head is released under CC BY 4.0. The base model Qwen/Qwen3.5-4B is under its own license.