tsmaitry/devica-toxicity-xlmr-large

DevicaAI Toxicity Detection — XLM-RoBERTa-Large

Multilingual toxicity classifier fine-tuned on XLM-RoBERTa-Large (559M params) for Indian language content moderation.

Supported Languages

Hindi, English, Bengali, Tamil, Telugu, Marathi, Kannada, Malayalam, Gujarati, Punjabi, Hinglish (code-mixed)

Performance

Metric	Score
Accuracy	~93-95%
F1 (weighted)	~93-95%
F1 (toxic)	~94%

Usage

from transformers import pipeline
classifier = pipeline("text-classification", model="tsmaitry/devica-toxicity-xlmr-large")
classifier("तुम बहुत अच्छे हो")   # → non-toxic
classifier("Saala kutta kamina")   # → toxic
classifier("நீ முட்டாள்")          # → toxic