Roblox’s open safety models reach 90.52 F1

Roblox open-sourced three safety models and a PII benchmark through ROOST, reporting that its updated PII classifier lifted F1 from 63.41 to 90.52. The company’s engineering report says the release also includes Sentinel v2 and a voice classifier; Sentinel swept 324 configurations in under three minutes, while its best ROC-AUC rose from 0.894 to 0.996. Moderation teams should file this as a reproducible starting point rather than a drop-in safety guarantee, adding the models to the layered review and evaluation perimeter described in the archive’s agent-evaluation governance analysis.