InferenceGuard ModernBERT Risk Classifier

Multi-task 4-head ModernBERT-base classifier that estimates per-comment inferential privacy risk across age, location, occupation, and education.

Trained on the profile-disjoint SynthPAI training split (RobinSta/SynthPAI, 210 profiles / 5,232 comments) using human inferability labels (reviews.human.<attr>.estimate) and dynamic positive class weighting (pos_weight BCE), followed by post-hoc temperature scaling (calibration_params.json) aligned to 0.30 / 0.60 risk bands (LOW < 0.30, MEDIUM 0.30โ€“0.59, HIGH >= 0.60).

Validation Metrics (45 held-out profiles, 1,416 comments)

Attribute F1 PR-AUC ROC-AUC ECE
Age 0.170 0.156 0.732 0.061
Location 0.544 0.525 0.898 0.017
Occupation 0.621 0.629 0.864 0.111
Education 0.600 0.580 0.868 0.051
Macro 0.484 0.473 0.840 0.060

Usage

from src.risk_model.predict import CalibratedRiskPredictor

predictor = CalibratedRiskPredictor.from_pretrained("AT707/inferenceguard-modernbert-risk")
report = predictor.predict("My co-op ends in December and I take the Green Line to campus.")

Ethics & Limitations

This model estimates attacker inferability risk from text; it does not guarantee anonymity or replace legal/compliance privacy controls. Trained on synthetic SynthPAI data (CC-BY-NC-SA-4.0).

Downloads last month
17
Safetensors
Model size
0.1B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for AT707/inferenceguard-modernbert-risk

Finetuned
(1,622)
this model