MHDC BERT-base

Fine-tuned from bert-base-uncased for 10-way classification of mental-health conditions and cyberbullying in social-media posts.

This repository is the retrained model: config, tokenizer, and weights from one fine-tuning run. It replaces an earlier upload that was only a converted checkpoint.

Held-out test set, 6,000 imbalanced posts: accuracy 0.876, macro F1 0.700.

Ajayi, Kachweka, Deku, and Aiken, PMLR 317:15-26, 2026. Best Paper, AAAI-26 AIMedHealth Bridge. https://proceedings.mlr.press/v317/ajayi26a.html

A screening aid with a person in the loop, not a diagnostic tool.

Labels

0 age_cb 1 anxiety 2 bipolar 3 ethnicity_cb 4 gender_cb 5 non_suicide 6 personality_disorder 7 religion_cb 8 stress 9 suicide

Use

from transformers import pipeline
clf = pipeline("text-classification", model="Jayi2424/mhdc-bert-base")
print(clf("I have not slept and I feel worthless."))