mentalbert-v5-deep-dive-v2

🧠 Deep Dive v2 — CSSRS-Anchored Binary Depression/Suicidal Re-Ranker

⚠️ CLINICAL SAFETY DISCLAIMER: This model is a research artefact for a mental health NLP thesis. It is NOT a certified clinical tool and must NOT be used for medical diagnosis, triage, or any safety-critical decision without qualified mental health professional oversight.

Architecture

BertTokenizerFast → CrisisEvidenceMentalBERT
                        ├─ BertModel (mental/mental-bert-base-uncased)
                        │    h_seq (B, L, 768)
                        ├─ h_cls   = h_seq[:, 0, :]               [CLS] pooling
                        └─ h_crisis = max-pool(h_seq, crisis_mask) lexical safety prior
                             (fallback: h_cls when no crisis tokens)
                        └─ [h_cls ; h_crisis] (1536)
                           → Linear(256) → GELU → Dropout → Linear(2) → P(Suicidal)

Crisis-evidence pooling injects a fixed lexical safety prior. When no crisis keywords are detected, the model collapses to standard [CLS] behaviour — no degradation on Depression-only inputs.

CSSRS-Anchored Training Data

V3's superior Sui→Dep safety (661) came from CSSRS-only Suicidal labels. V5 source unification degraded this signal. Deep Dive v2 restores V3-quality training signal without distorting the test distribution.

Threshold Calibration

Threshold calibrated against Sui→Dep error budget (≤200 on val), NOT F1 macro.

• Sweep: t ∈ [0.20, 0.80] step 0.05
• Primary: Sui→Dep ≤ 200 → maximise F1 macro
• Fallback: minimise Sui→Dep with Dep→Sui ≤ 500

Chosen threshold: 0.25

Test Metrics (all-source frozen test set)

Intended Use

Must NOT be used standalone. Invoked by Quick Vibe (itsLu/mentalbert-v5-source-aware) only when its top prediction ∈ {Depression, Suicidal} with a low-confidence margin, or when it abstains. No coverage for Normal / Anxiety / Stress / Bipolar / Personality Disorder / Directed Aggression.

Quick Start

from inference import DeepDiveV2
clf    = DeepDiveV2.from_hub("itsLu/mentalbert-v5-deep-dive-v2", device="cpu")
result = clf.predict("I don't want to be here anymore.")
# {'label': 'Suicidal', 'p_suicidal': 0.913, 'crisis_evidence_found': True,
#   'crisis_tokens_matched': ['want to die']}

Citation

Part of: Multi-Source Mental Health Text Classification with CSSRS-Anchored Label Curation, Thesis 2026. Dataset: mohamedasem318/mental-health-dataset-extended-v5.