Moderation classifier demo

Portfolio demo, not a production moderation system. DistilBERT fine-tuned on the public Civil Comments data (CC0) by Malav Naik. It runs in your browser: nothing you type is sent anywhere. Known limit: harmless comments that mention some identities are flagged 2–4× more often than average.

Send-to-review threshold:

Loading model (about 70 MB, once)…

Lower the threshold to catch more toxic comments, and wrongly flag more harmless ones. Auto-action needs toxicity ≥ (90% precision on held-out data).