itsLu/mentalbert-v5-flat-8class

MentalBERT V5 — Flat 8-Class Mental Health Classifier

Single-pass MentalBERT fine-tuned on the V5 mental-health dataset. Predicts one of 8 classes:

Anxiety, Bipolar, Depression, Directed Aggression, Normal, Personality Disorder, Stress, Suicidal.

Test Set Results (V5 stratified 70/10/20, random_state=42)

Metric	Value
Accuracy	82.84%
F1 macro	0.8350
F1 weighted	0.8280
Sui→Dep (missed crises)	516
Total Dep↔Sui bleed	1249
ROC AUC (macro)	0.9638

Quick Start (Python)

from transformers import pipeline

clf = pipeline("text-classification", model="<YOUR_USERNAME>/mentalbert-v5-flat-8class")
result = clf("I haven't slept in days, I feel like everything is falling apart.")
print(result)  # [{'label': 'Stress', 'score': 0.87}]

API Call (HF Inference Endpoint)

import requests
HF_TOKEN = "hf_..."
URL = "https://api-inference.huggingface.co/models/<YOUR_USERNAME>/mentalbert-v5-flat-8class"
headers = {"Authorization": f"Bearer {HF_TOKEN}"}
r = requests.post(URL, headers=headers, json={"inputs": "I want to end it all."})
print(r.json())  # [{'label': 'Suicidal', 'score': 0.91}, ...]

For top-k probabilities over all classes, pass {"inputs": text, "parameters": {"top_k": 8}}.

Limitations

• This is a screening signal, not a clinical diagnosis. Use only as one input among many.
• Sui→Dep (516) errors are missed crisis cases. Pair with a safety threshold or with the hierarchical companion model (mentalbert-v5-hierarchical-longformer) for safety-critical applications.
• Trained on Reddit-style English text; out-of-distribution domains (clinical notes, formal prose) may degrade.

Training Details

• Backbone: mental/mental-bert-base-uncased
• MAX_LEN: 128, batch=32, LR=2e-05, epochs=4, label_smoothing=0.05
• Imbalance: WeightedRandomSampler + class-weighted CrossEntropy (cap=3.0)
• Hardware: NVIDIA T4 (Kaggle)