XLM_2_MIXED_DISTIL

Bangla Hate Speech Detection Model

This model is fine-tuned for binary hate speech detection in Bangla text.

Model Description

• Base Model: AnnNaserNabil/BaseModelXLMRoberta
• Task: Binary Classification (Hate Speech vs Non-Hate Speech)
• Language: Bangla (Bengali)
• Training Method: Baseline → Knowledge Distillation

Training Details

Training Hyperparameters

• Batch Size: 32
• Learning Rate: 2e-05
• Epochs: 15
• Max Sequence Length: 128
• Dropout: 0.1
• Weight Decay: 0.01
• Warmup Ratio: 0.1

Training Data

• K-Fold Cross-Validation: 5 folds
• Stratification: binary

Performance

Add your metrics here after training

Usage

from transformers import AutoModel, AutoTokenizer
import torch
import torch.nn as nn
import json

# Load model components
encoder = AutoModel.from_pretrained("path/to/model")

with open("path/to/model/classifier_config.json", 'r') as f:
    c_config = json.load(f)

classifier = nn.Sequential(
    nn.Linear(c_config['hidden_size'], 256),
    nn.ReLU(),
    nn.Dropout(0.1),
    nn.Linear(256, c_config['num_labels'])
)
classifier.load_state_dict(torch.load("path/to/model/classifier.pt"))

tokenizer = AutoTokenizer.from_pretrained("path/to/model")

# Predict
def predict(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    with torch.no_grad():
        outputs = encoder(**inputs)
        cls_embedding = outputs.last_hidden_state[:, 0, :]
        logits = classifier(cls_embedding)
        prob = torch.sigmoid(logits).item()
    return prob

text = "আপনার বাংলা টেক্সট এখানে"
prob = predict(text)
print(f"Hate Speech Probability: {prob:.4f}")

Citation

If you use this model, please cite:

@misc{bangla-hate-speech-model,
  author = {Nabil},
  title = {Bangla Hate Speech Detection Model},
  year = {2026},
  publisher = {HuggingFace},
}

License

MIT License