urdu-spam-classifier

Urdu Spam Detection Model

Description

This model classifies Urdu text into:

• 0 → Not Spam
• 1 → Spam

It is designed for AI-powered emergency helpline systems (e.g., 1122/911) to filter prank or irrelevant calls in real time.

---

Model Details

• Base Model: xlm-roberta-base
• Task: Binary Text Classification
• Language: Urdu
• Framework: Hugging Face Transformers

---

Training

• Dataset: hamza-amin/urdu-spam-dataset
• Epochs: 3
• Batch size: 16
• Learning rate: 2e-5

---

Evaluation Results

---

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model_name = "hamza-amin/urdu-spam-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
def predict(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=1)
    return probs.tolist()
print(predict("یہ ایک ایمرجنسی ہے فوراً مدد کریں"))
print(predict("ہیلو بس مذاق کر رہا تھا"))

---

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Use Cases

• Emergency call filtering
• Spam detection in voice agents
• Urdu NLP classification tasks

---

Limitations

• Trained on synthetic data

• May struggle with:

  • slang / mixed Urdu-English
  • very short inputs
  • noisy speech-to-text output

---

Future Work

• Add urgency classification (multi-class)
• Train on real call data
• Optimize for real-time inference (ONNX / quantization)

---

License

MIT