PNT Score Predictor

Paste a discourse transcript (Cinderella story retelling, or PBJ procedure description) and get predicted Philadelphia Naming Test (PNT) scores — correct responses plus 7 error subtypes.

Everything runs in your browser. The transcript you paste is never uploaded anywhere — feature extraction, the sentence embedding, and the trained regression models all run client-side (via transformers.js and onnxruntime-web). Only static model files are fetched (from this site, plus the public embedding-model weights on first load).

Research use only. Not clinically validated. Does not replace formal PNT administration.

How it was built

The prediction models (Ridge / RandomForest / XGBoost, one per PNT score type) were trained in Python on paired discourse-transcript + PNT-score data, then converted to ONNX and plain JSON so they run natively in the browser with no server. See the parent project's train_final_models.py and web/dev/export_web_models.py for the training/export pipeline.