mentalbert-v6-hierarchical

MentalBERT V6 Hierarchical — 5-Stage Cascade

A 5-stage hierarchical mental-health text classifier covering 8 classes: Normal, Depression, Suicidal, Anxiety, Stress, Bipolar, Personality Disorder, Directed Aggression.

This is the Deep Dive tier of the Vibecheck project (Helwan University). The Quick Vibe counterpart is itsLu/mentalbert-v6-flat.

Architecture

text → Stage 0  (Cardiff RoBERTa, DA binary)
           ↓
       Stage 1A (MentalBERT, Suicidal binary)
           ↓
       Stage 1B (MentalBERT, Normal vs Distress)
           ↓
       Stage 2  (MentalBERT, 5-class Anxiety/Bipolar/Dep/PD/Stress)
           ↓
       Stage 3  (Longformer, Depression vs Suicidal re-scorer)

Each stage is fine-tuned on the full V6 dataset, carved per stage. Routing thresholds are calibrated on the validation set via a joint t1A × t3 grid sweep producing two operating points:

• BALANCED — maximises F1_macro − 0.30·Sui_miss_rate − 0.30·Sui_FP_rate.
• SAFETY — constrains Sui→Dep ≤ 150 on val, then minimises Dep→Sui.

Usage (HF Inference Endpoint)

import requests
r = requests.post(ENDPOINT_URL,
    headers={"Authorization": f"Bearer {HF_TOKEN}"},
    json={"inputs": "I don't see the point anymore.", "mode": "safety"})
print(r.json())
# -> {"label": "Suicidal", "exit_stage": "stage1a", "mode": "safety",
#     "stage_probs": {...}}

mode is "balanced" (default) or "safety".

Files

• stage0/ — Cardiff RoBERTa, DA gate
• stage1a/ — MentalBERT, Suicidal gate
• stage1b/ — MentalBERT, Normal/Distress
• stage2/ — MentalBERT, 5-class distress
• stage3/ — Longformer, Dep/Sui re-scorer
• config.json — thresholds, class order, metrics
• handler.py — HF Inference Endpoints handler with mode switching

Training data

mohamedasem318/mental-health-dataset-extended-v6 — ~89k rows from 7 sources (cssrs, jigsaw17, civil_comments, kaggle_bpd, kaggle, huggingface, swmh). V6 replaces V5's OLID-sourced Directed Aggression class with Jigsaw 2017 + Civil Comments threat-labelled data — the change targets real-world detection of explicit violent intent (e.g., "I want to kill my friend"), which OLID's Twitter-insult distribution did not cover. Stratified 70/10/20 split, random_state=42. Per-source reliability constants are recorded in config.json but were not applied to the loss (plain class-weighted CE only — a prior attempt with a per-sample-weighted wrapper destabilised Stage 0 training under fp16/AMP).

Citation

If you use this model, please cite the Vibecheck project (Helwan University final-year project, 2026).