---
language:
- ur
library_name: transformers
pipeline_tag: text-classification
tags:
- roberta
- urdu
- hate-speech
- sequence-classification
- pytorch
- smote
- tf-idf
license: other  # inherit/align with base model's license
datasets:
- Adnan855570/urdu-hate-speech
---

## Urdu RoBERTa Hate Speech Classifier (Balanced)

- **Base model**: `urduhack/roberta-urdu-small`
- **Task**: Binary text classification (hate vs. not_hate)
- **Language**: Urdu (ur)
- **Labels**
  - 0 → `not_hate`
  - 1 → `hate`

This model fine-tunes a small RoBERTa for Urdu hate-speech detection. Class imbalance was addressed by oversampling with SMOTE at the feature level (TF–IDF) prior to tokenization-based training.

### Training data and preprocessing
- Source dataset: `Adnan855570/urdu-hate-speech` (Excel files: `preprocessed_combined_file (1).xlsx`, `Urdu_Hate_Speech.xlsx`)
- Columns used in notebook: `Tweet` (text), `Tag` (label in {0,1})
- Steps:
  - TF–IDF featurization (max_features=10000)
  - SMOTE oversampling (random_state=42) to balance classes
  - Train/test split: 80/20 (random_state=42)
  - Tokenization: `AutoTokenizer.from_pretrained("urduhack/roberta-urdu-small")` with `truncation=True`, `padding=True`

### Training setup
- Model: `AutoModelForSequenceClassification` with `num_labels=2`
- Device: GPU if available
- Hyperparameters:
  - epochs: 3
  - per_device_train_batch_size: 8
  - per_device_eval_batch_size: 8
  - warmup_steps: 500
  - weight_decay: 0.01
  - evaluation_strategy: epoch
  - save_strategy: epoch
  - load_best_model_at_end: true
- Metrics:
  - Accuracy, Precision, Recall, F1 (binary)

### Evaluation results (test split)
- accuracy: 0.7891
- f1: 0.7854
- precision: 0.8208
- recall: 0.7529

Note: Results derive from the balanced (SMOTE) dataset and the 80/20 split used in the notebook.

### How to use (Transformers)

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

MODEL_ID = "Adnan855570/urdu-roberta-hate"  # replace if different
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID).eval()

id2label = model.config.id2label or {"0":"not_hate","1":"hate"}

def predict(text: str):
    enc = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    with torch.no_grad():
        logits = model(**enc).logits
        probs = logits.softmax(dim=-1).squeeze().tolist()
        pred = int(logits.argmax(dim=-1).item())
    return {"label_id": pred, "label": id2label.get(str(pred), str(pred)),
            "scores": {"not_hate": probs[0], "hate": probs[1]}}

print(predict("یہ نفرت انگیز ہے یا نہیں؟"))
```

Or with a pipeline:

```python
from transformers import pipeline
clf = pipeline("text-classification", model="Adnan855570/urdu-roberta-hate", top_k=None)
print(clf("یہ نفرت انگیز ہے یا نہیں؟"))
```

### Inference API

- cURL
```bash
curl -X POST -H "Authorization: Bearer $HF_TOKEN" -H "Content-Type: application/json" \
  -d '{"inputs":"یہ نفرت انگیز ہے یا نہیں؟"}' \
  https://api-inference.huggingface.co/models/Adnan855570/urdu-roberta-hate
```

- Python
```python
import os, requests
API_URL = "https://api-inference.huggingface.co/models/Adnan855570/urdu-roberta-hate"
HEADERS = {"Authorization": f"Bearer {os.environ.get('HF_TOKEN','')}"}
print(requests.post(API_URL, headers=HEADERS, json={"inputs":"..."}, timeout=30).json())
```

### Intended uses and limitations
- Intended:
  - Flagging potentially hateful Urdu content
  - Assisting human moderation and research
- Limitations:
  - May misclassify satire, reclaimed slurs, or code-mixed content
  - Domain shift sensitivity (platform/community/topic)
- Risks:
  - False positives/negatives; do not use as the sole basis for punitive actions
- Recommendation:
  - Use with human-in-the-loop; periodically audit outcomes and bias

### Label mapping
Ensure the config includes:
- `id2label = {"0":"not_hate","1":"hate"}`
- `label2id = {"not_hate":0,"hate":1}`

### Reproducibility notes
- SMOTE and split seeds: `random_state=42`
- Tokenization: truncation and padding enabled (no explicit max_length set in notebook)
- Hardware: single GPU (e.g., Colab)

### License
- The model derivation should comply with the base model’s license (`urduhack/roberta-urdu-small`). Set a compatible license here once confirmed.

### Citation
```bibtex
@misc{urdu_roberta_hate_balanced_2025,
  title  = {Urdu RoBERTa Hate Speech Classifier (Balanced)},
  author = {Adnan},
  year   = {2025},
  howpublished = {\url{https://huggingface.co/Adnan855570/urdu-roberta-hate}}
}
```

### Acknowledgements
- Base: `urduhack/roberta-urdu-small`
- Libraries: 🤗 Transformers, Datasets, PyTorch
- Oversampling: SMOTE (imblearn)