detoxify

Why we included this project

Running a comments section, forum, or social feed often means deciding quickly which posts need a human look. Detoxify gives you a trained baseline for that: models built on the three Jigsaw Toxic Comment Challenge datasets, covering the standard categories those competitions targeted. Inference is straightforward from Python, and the lightweight ALBERT variants make it easy to run when speed or memory matter. If your content skews multilingual or you want to reduce bias, the corresponding model weights are available, and the codebase is set up so you can fine-tune further. It's a practical drop-in filter for moderation, and also a clean example of structuring a sentence-classification project with PyTorch Lightning and Transformers.