---
language:
- en
- hi
license: mit
tags:
- text-classification
- hate-speech-detection
- xlm-roberta
- multilingual
datasets:
- hasoc2019
metrics:
- accuracy
- f1
pipeline_tag: text-classification
widget:
- text: I love everyone in this community!
  example_title: Positive Example
- text: This person is terrible and should be banned
  example_title: Negative Example
---

# Hate Speech Detector (XLM-RoBERTa)

Multilingual hate speech detection model fine-tuned on HASOC 2019 dataset.

## Model Description

This model detects hate speech in English and Hindi text using XLM-RoBERTa base as the backbone.

**Languages:** English, Hindi  
**Task:** Binary Text Classification (Hate Speech / Not Hate Speech)  
**Base Model:** xlm-roberta-base

## Intended Uses

- Content moderation
- Social media monitoring  
- Research purposes

## How to Use

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

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("archich/hate-speech-detector")
model = AutoModelForSequenceClassification.from_pretrained("archich/hate-speech-detector")

# Example text
text = "Your text here"

# Tokenize
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=256)

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=1)
    prediction = torch.argmax(probs, dim=1).item()

labels = ["NOT_HATE_SPEECH", "HATE_SPEECH"]
print(f"Prediction: {labels[prediction]} ({probs[0][prediction].item():.2%} confidence)")
```

## Training Data

Trained on HASOC 2019 (Hate Speech and Offensive Content Identification) dataset containing:
- Hindi posts from social media
- English posts from social media

## Label Mapping

- `0`: NOT_HATE_SPEECH - Normal, non-offensive content
- `1`: HATE_SPEECH - Hateful or offensive content (HOF)

## Limitations & Ethical Considerations

⚠️ **Important Notice:**

- This model is intended to **assist** human moderators, not replace them
- May contain biases from training data
- Context and cultural nuances are important - manual review recommended
- False positives are possible
- Should not be the sole decision-maker for content removal

## Performance

Training details and metrics available in model files.

## Citation

If you use this model, please cite:

```
@misc{hate-speech-detector,
  author = {archich},
  title = {Multilingual Hate Speech Detector},
  year = {2024},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/archich/hate-speech-detector}}
}
```