pshashid/xlmr-toxicity-multilingual

XLM-RoBERTa Toxicity Detection (Multilingual)

Overview

This model detects toxic comments across 14 languages using a single, unified model.

Supported Languages

German, Spanish, French, Italian, Russian, Chinese, Japanese, Arabic, Hebrew, Amharic, Tatar, Ukrainian, Hindi

Model Details

• Architecture: XLM-RoBERTa-Large (384M parameters)
• Task: Binary toxicity classification
• Training Data: TextDetox Multilingual Toxicity Dataset

Per-Language Performance (F1-Score)

• French: 0.9931
• Russian: 0.9877
• Hindi: 0.9403
• Ukrainian: 0.9401
• Japanese: 0.8917
• Spanish: 0.8846
• Italian: 0.8834
• German: 0.8823
• Tatar: 0.8681
• Arabic: 0.8310
• Amharic: 0.8267
• Chinese: 0.8150
• Hebrew: 0.7829

Usage

import torch
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download
from model_utils import MultiTaskXLMR

# Load model
model_path = hf_hub_download("{repo_id_multilingual}", "pytorch_model.bin")
model = MultiTaskXLMR(model_name="xlm-roberta-large", num_intents=5)
model.load_state_dict(torch.load(model_path))
model.eval()

# Inference in any supported language
tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-large")
text = "Tu es un idiot!"  # French
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)

with torch.no_grad():
    tox_logits, _ = model(inputs['input_ids'], inputs['attention_mask'])
    tox_prob = torch.sigmoid(tox_logits).item()
    print(f"Toxicity: {tox_prob:.4f}")

License

MIT