Laya for toxic comment moderation (Jigsaw)
Answers 6 yes/no questions per comment (toxic, severe_toxic, obscene, threat, insult, identity_hate) with calibrated probabilities. Fine-tuned on 12k balanced Jigsaw comments, then weight-merged with the base model (WiSE-FT, 50% fine-tuned; mix chosen on validation only).
| Model | Mean ROC-AUC, full Jigsaw test set (63,978 comments) |
|---|---|
| This model | 0.9853 |
| Detoxify original | 0.9863 |
| TF-IDF + logistic regression | 0.9786 |
ECE 0.0223 · ~18 comments/sec on one T4 · $0/request self-hosted.