unitaryai/detoxify

Provides three distinct model variants—`original`, `unbiased`, and `multilingual`—each optimized for different toxicity detection scenarios, with lightweight ALBERT-based alternatives for resource-constrained deployments. Leverages transformer-based architectures with bias-aware training on aggregated annotator judgments, supporting multi-label classification across toxicity subtypes (obscenity, threats, identity attacks, etc.) and identity mentions. Exposes predictions via a simple Python API returning per-category confidence scores and supports inference across seven languages with per-language performance metrics.

1,202 stars and 94,691 monthly downloads. Used by 4 other packages. Actively maintained with 2 commits in the last 30 days. Available on PyPI.