This is a Domain-Adaptive Pre-Trained (DAPT) model for detecting depression severity in Bangla social media text. A robust multilingual model capable of handling mixed-language (Code-Mixed) scenarios.

Model Details

Base Architecture: google-bert/bert-base-multilingual-cased
Training Method: Domain-Adaptive Pre-Training (DAPT) on bangla mental health corpora, followed by Fine-Tuning.
Task: Multi-class Classification (4 classes).
Language: Bengali (Bangla).

Label Mapping

The model outputs one of the following classes:

0: Minimum/None
1: Mild
2: Moderate
3: Severe

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load the model
model_name = "SrothJr/banglamentalBERT-mBERT-dapt"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Inference
text = "আপনার বাংলা টেক্সট এখানে (Your Bengali text here)"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    
predictions = torch.argmax(outputs.logits, dim=-1)
labels = ["Minimum", "Mild", "Moderate", "Severe"]
print(f"Prediction: {labels[predictions.item()]}")

Citation

If you use this mode, please cite our research.