Fine-tuned DistilBERT for binary depression classification from social media text (Twitter + Reddit).
Trained as part of the ExplainDepression Pipeline — a research system for explainable and fair depression detection using SHAP attributions, Depression Explanation Score (DES), and Explanation Consistency Index across Demographics (ECID).
━━━━━━━━━━━━━━━━━━━━━━━━━━━
PERFORMANCE Accuracy → 96.17% F1 Score → 0.9617 AUC-ROC → 0.9937 ECID → 0.2063 (bias detected in 3 clinical categories)
━━━━━━━━━━━━━━━━━━━━━━━━━━━
TRAINING DETAILS Base Model → distilbert-base-uncased Epochs → 3 Batch Size → 64 Learning Rate → 3e-5 Max Length → 96 tokens Train Size → 11,200 posts Platform → Kaggle GPU T4
━━━━━━━━━━━━━━━━━━━━━━━━━━━
LABELS LABEL_0 → Not Depressed LABEL_1 → Depressed
━━━━━━━━━━━━━━━━━━━━━━━━━━━
HOW TO USE
from transformers import pipeline
clf = pipeline( "text-classification", model="mdsajjadullah/explainDepression-distilbert", return_all_scores=True )
result = clf("I feel hopeless and completely alone.")
LABEL_1 score = depression probability
━━━━━━━━━━━━━━━━━━━━━━━━━━━
INTENDED FOR RESEARCH ONLY. Not for clinical diagnosis or deployment without professional mental health oversight.