itsLu/mentalbert-v5-quick-vibe-v2

🧠 MentalBERT Quick Vibe v2

Calibrated, safety-aware mental health text classifier with abstention.
Post-hoc upgrade to itsLu/mentalbert-v5-source-aware — encoder weights are unchanged.

Model Description

Quick Vibe v2 wraps the cooperative source-aware MentalBERT encoder with three post-hoc layers:

1. Per-class temperature scaling — fit on validation NLL via LBFGS; reduces overconfidence and improves calibration (ECE).
2. Asymmetric Dep/Sui safety margin — when Depression and Suicidal are the top two predictions with a margin < threshold, the prediction is flipped to Suicidal (the safer clinical alarm).
3. Calibrated abstention — when top-1 calibrated confidence < threshold, the model returns ABSTAIN rather than a low-confidence guess.

Architecture

SourceAwareMentalBERT (frozen):
  bert: mental/mental-bert-base-uncased
  cls_head: nn.Linear(768, 8)   ← loaded from itsLu/mentalbert-v5-source-aware

Post-hoc stack (this repo):
  Per-class temperature T ∈ ℝ⁸  ← calibration.pt
  Asymmetric safety margin       ← decision_thresholds.json
  Abstention threshold           ← decision_thresholds.json

Dataset

Trained and calibrated on mohamedasem318/mental-health-dataset-extended-v5
(stratified 70/10/20 split, random_state=42).

Performance

Run	Coverage	Acc (covered)	F1 Macro	Sui→Dep	Dep→Sui	Total Bleed
v1 (deployed)	100.0%	83.23%	0.8381	544	715	1,259
v2 (this)	89.8%	87.48%	0.8763	354	553	907

Intended Use

Intended for mental health screening as a second-tier triage component
(Vibecheck deployment). Not a diagnostic instrument.
All outputs require human clinical review before action.

Out of Scope

• NOT a diagnostic instrument. This model cannot diagnose mental health conditions.
• NOT a replacement for clinical judgement. All ABSTAIN and Suicidal predictions must be reviewed by a qualified professional.
• NOT validated for production without a prospective clinical study.

⚠️ Clinical Safety Disclaimer

This model is a research prototype. Predictions, including Suicidal classifications,
are probabilistic and will contain errors. Do not use as the sole basis for
clinical decisions. Always involve a qualified mental health professional.

Usage

from huggingface_hub import hf_hub_download
import runpy, sys

# Download inference.py and run it in your environment
path = hf_hub_download(repo_id='itsLu/mentalbert-v5-quick-vibe-v2', filename='inference.py')

# Or: import directly once the file is on your path
# from inference import predict
# result = predict("I feel hopeless and empty.")
# print(result)
# {
#   'label': 'Depression',
#   'confidence': 0.72,
#   'all_probs': {...},
# #   'abstained': False, 'safety_flipped': False
# }

Files

File	Description
calibration.pt	Per-class temperature tensors
decision_thresholds.json	abstain_thresh, dep_sui_margin, metrics
inference.py	Self-contained load + predict script

Citation

@misc{mentalbert-quickvibe-v2,
  author  = {Mohamed Asem},
  title   = {MentalBERT Quick Vibe v2: Calibrated Safety-Aware Mental Health Classifier},
  year    = {2025},
  url     = {https://huggingface.co/itsLu/mentalbert-v5-quick-vibe-v2}
}