You can deploy this model and start earning money today!

Sherazahmad634/
mental-health-support-tier-classifier

Name: mental-health-support-tier-classifier
Availability: InStock

sentence-transformers

mental-health

sentence-transformers/all-MiniLM-L6-v2

finetune

mit

Views

No views yet

Mental-Health Support Tier Classifier

A logistic-regression linear probe on top of frozen sentence-transformers/all-MiniLM-L6-v2 embeddings. Given a short user message, it predicts one of four support tiers: self_help, peer_support, professional, crisis.

How it works

text ─► all-MiniLM-L6-v2 (frozen) ─► 384-dim vector ─► Logistic Regression ─► tier

The classifier is shipped as a single joblib bundle that contains:

backend — "st"
embedder — "sentence-transformers/all-MiniLM-L6-v2"
classifier — fitted sklearn.linear_model.LogisticRegression
label_to_id / id_to_label / label_names
trained_at — ISO timestamp

Usage

python

import joblib, numpy as np
from sentence_transformers import SentenceTransformer

bundle = joblib.load("classifier.joblib")
embedder = SentenceTransformer(bundle["embedder"])
clf = bundle["classifier"]

text = "I've felt numb for almost six months and exercise doesn't help anymore."
vec = embedder.encode([text], normalize_embeddings=True)
proba = clf.predict_proba(vec)[0]
for label, p in zip(bundle["label_names"], proba):
    print(f"{label:14s}  {p:.3f}")

Training

Base model: sentence-transformers/all-MiniLM-L6-v2 (frozen, no fine-tuning).
Classifier: LogisticRegression(max_iter=2000, C=4.0, class_weight="balanced").
Data: <your-username>/mental-health-support-tier (synthetic, ~1000 rows).
Selection: macro-F1 on a 200-row stratified test split, compared against calibrated linear SVM and a small MLP.

Evaluation

Metrics on the held-out test split and on a 20-row hand-written adversarial set are included in the repo (metrics.json, adversarial_metrics.json). The confusion matrix is in confusion_matrix.png.

Intended use

Educational demonstration of the embeddings → linear-probe → demo workflow. The companion Space at <your-username>/mental-health-support-tier-demo always surfaces crisis resources regardless of the model's prediction.

Limitations

Trained on synthetic, templated data. It will degrade on user text that is far from the training distribution (slang, code-switching, sarcasm).
The label boundaries are author-defined, not clinically validated.
The model has no notion of who is talking — context, history, user intent, cultural framing, etc. are all out of scope.
Do not use it for real-world triage decisions.

License

MIT.

Sherazahmad634/mental-health-support-tier-classifier