Open Source Context-Aware PII Classifier (corp.roblox.com)

Roblox has open-sourced its context-aware PII Classifier — an AI model built to detect not just explicit personally identifiable information (PII) but conversational intent to share or solicit PII. The company says the released model (part of its safety toolkit) has shown strong production results: 98% recall at 1% false-positive rate on an internal test set and a 94% F1 on over 47,000 labeled real-world samples, outperforming other safety solutions like LlamaGuard v3 8B and Piiranha NER by large margins. Roblox framed the release as a practical response to adversarial language (character obfuscation, implicit platform references, slang) that conventional token/NER-based detectors miss, and emphasized ongoing adaptation through human and AI red‑teaming. Technically, the classifier is an encode-only transformer fine-tuned from XLM-RoBERTa-Large, served with a split architecture (tokenizer and pre/post-processing on CPU; transformer on GPU) with dynamic batching on Triton to sustain peaks of ~200k queries/sec at <100ms P90 latency. Training combined targeted samplers (uncertainty sampling and consecutive-block sampling) with AI-generated synthetic data to expose rare or emerging bypass patterns, plus iterative red-teaming to close gaps. For the AI/ML community this release offers a production-proven, context-aware PII detection approach and a concrete example of combining sampling strategies, synthetic augmentation, and infrastructure choices to balance robustness and low-latency scale.