bangla-cyberbully-google-bert-base-multilingual-cased

Bangla Cyberbullying Detection Model

This model is fine-tuned for multi-label classification to detect cyberbullying in Bangla text.

Model Details

• Base Model: google-bert/bert-base-multilingual-cased
• Task: Multi-label text classification
• Labels: bully, sexual, religious, threat, spam
• Number of Labels: 5
• Classifier Hidden Size: 256
• Dropout: 0.1

Usage

Installation

pip install torch transformers

Loading and Inference

from model import TransformerMultiLabelClassifier
from transformers import AutoTokenizer
import torch

# Load the model
model = TransformerMultiLabelClassifier.from_pretrained("path/to/saved/model")
tokenizer = AutoTokenizer.from_pretrained("path/to/saved/model")

# Prepare input
text = "আপনার বাংলা টেক্সট এখানে"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)

# Get predictions
outputs = model.predict(inputs['input_ids'], inputs['attention_mask'])
    
probabilities = outputs['probabilities'][0]
predictions = outputs['predictions'][0]

labels = ['bully', 'sexual', 'religious', 'threat', 'spam']
for label, prob, pred in zip(labels, probabilities, predictions):
    status = "✓ Detected" if pred else "✗ Not detected"
    print(f"{label}: {prob:.4f} ({status})")

Using with Pipeline (Alternative)

# For batch inference
texts = ["টেক্সট ১", "টেক্সট ২", "টেক্সট ৩"]
inputs = tokenizer(texts, return_tensors="pt", truncation=True, padding=True, max_length=128)
outputs = model.predict(inputs['input_ids'], inputs['attention_mask'])

Labels

Training

This model was trained using:

• K-fold cross-validation with multi-label stratification
• AdamW optimizer with linear warmup
• Mixed precision training (AMP)
• Early stopping based on weighted F1 score

Citation

If you use this model, please cite:

@misc{bangla-cyberbullying-detection,
  author = {Your Name},
  title = {Bangla Cyberbullying Detection Model},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/your-username/your-model}
}

Limitations

• Trained specifically on Bangla text
• Performance may vary on out-of-domain text
• Multi-label threshold of 0.5 used by default (can be adjusted)
• May not generalize well to code-mixed text (Bangla + English)