---
language:
- en
- ru
license: mit
tags:
- text-classification
- hate-speech-detection
- toxicity
- xlm-roberta
- bilingual
datasets:
- davidson-hate-speech
- russian-toxic-comments
metrics:
- f1
- accuracy
model-index:
- name: Hate-Speech-Detector-XLM-RoBERTa
  results:
  - task:
      type: text-classification
      name: Hate Speech Detection
    metrics:
    - type: f1
      value: 0.925
      name: Macro F1
    - type: accuracy
      value: 0.933
      name: Accuracy
base_model:
- FacebookAI/xlm-roberta-base
pipeline_tag: text-classification
---

# 🛡️ XLM-RoBERTa Hate Speech Detector (EN/RU)

Multilingual toxic comment classification model fine-tuned on English and Russian datasets.

## Model Description

- **Base Model**: `xlm-roberta-base`
- **Languages**: English, Russian
- **Task**: Binary text classification (non-toxic / toxic)
- **Training**: Fine-tuned on Davidson English + Russian Toxic Comments datasets

## Performance

### Overall Metrics
- **Macro F1**: 0.925
- **Accuracy**: 0.933

### Language-Specific Performance

**English:**
- Macro F1: 0.900
- Non-toxic F1: 0.831
- Toxic F1: 0.968
- FPR: 0.220
- FNR: 0.019

**Russian:**  
- Macro F1: 0.900
- Non-toxic F1: 0.930
- Toxic F1: 0.871
- FPR: 0.094
- FNR: 0.082

## Usage

```python
import torch
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(
    repo_id="Anchar21/Hate-Speech-Detector-XLM-RoBERTa",
    filename="model.pt"
)

# Load model architecture (you need BertClassifier class)
from your_module import BertClassifier

model = BertClassifier(
    model_name="xlm-roberta-base",
    num_labels=2,
    dropout=0.1
)

# Load weights
checkpoint = torch.load(model_path, map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")

# Inference
text = "Your text here"
encoding = tokenizer(
    text,
    max_length=128,
    padding='max_length',
    truncation=True,
    return_tensors='pt'
)

with torch.no_grad():
    logits = model(encoding['input_ids'], encoding['attention_mask'])
    probs = torch.softmax(logits, dim=1)
    pred = torch.argmax(logits, dim=1)

print(f"Prediction: {'non-toxic', 'toxic'}[pred.item()]")
print(f"Confidence: {probs[0][pred].item():.3f}")
```

## Training Details

- **Learning Rate**: 1e-5
- **Batch Size**: 16
- **Epochs**: 3
- **Class Weights**: True
- **Max Sequence Length**: 128

## Limitations

- English texts show higher false positive rate (~22%) - model is aggressive on borderline cases
- Trained on specific datasets - may not generalize to all domains
- Binary classification only (no severity levels)

## Citation

```bibtex
@misc{xlm-roberta-hate-speech-en-ru,
  author = {Anchar21},
  title = {XLM-RoBERTa Hate Speech Detector},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Anchar21/Hate-Speech-Detector-XLM-RoBERTa}}
}
```

## License

MIT