Detoxify
Open models scoring text for toxicity, threats and identity attacks.
About Detoxify
Detoxify provides trained models and code that predict toxic comments across all three Jigsaw Toxic Comment Challenges. Laura Hanu at Unitary built Detoxify. The page states that the intended use is research, fine-tuning on carefully constructed datasets, and aiding content moderators in flagging harmful content.
The multilingual model is stated to cover seven languages: English, French, Spanish, Italian, Portuguese, Turkish and Russian.
Drafted from github.com · not yet confirmed by Unitary
How it works
Detoxify is a set of open PyTorch models from Unitary trained on the Jigsaw toxic comment datasets. There are three: original (toxic, severe toxic, obscene, threat, insult, identity hate), unbiased (trained to reduce identity term bias) and multilingual (seven languages). Install with pip, call predict, get scores.
It's the simplest way to run a Perspective style classifier on your own hardware, with the same limits: it reflects the training data's era and English centric labels.
Reviews
Be the first to review it
AnonymousKnow a team weighing Detoxify? Send them this page.