Roblox PII Classifier
RobloxRoblox PII Classifier flags chat messages that ask for or share personal information, such as phone numbers, social media handles and credentials, and messages that try to move a user to another platform. Version 2 reads each message together with the surrounding multi-user conversation, which helps it catch bypass attempts like phonetic spelling and implicit references. Roblox built it by fine-tuning the 560M-parameter XLM-RoBERTa-Large encoder for its own chat moderation.
A ROOST Model Community partner model. Version 1 stays on Hugging Face as its own repository and is counted here as an earlier release of the same product.
Openness
3 high confidence- weights
- open(Apache-2.0 safetensors and ONNX on HF Roblox/roblox-pii-classifier-v2, with no access gate)
- data
- described(the cards and Roblox's announcement describe expert-labeled, anonymized Roblox chat plus AI-generated conversations. Neither is released, and the Roblox dataset on the Hub is an evaluation-only benchmark)
- code
- partial(an inference script and requirements file only, no training code)
- license
- Apache-2.0(OSI)
- prior-release
- distributed(Roblox PII Classifier v1 (Roblox/roblox-pii-classifier) stays downloadable under Apache-2.0)
The v2 weights are published under Apache 2.0 with no access gate, along with an inference script. Roblox describes the labeled chat and synthetic conversations it trained on but releases neither them nor its training code, so this is an open-weights model rather than a fully open one.
- https://about.roblox.com/newsroom/2025/11/open-sourcing-roblox-pii-classifier-ai-pii-detection-chat recorded 2026-10-19
Roblox's announcement of November 4, 2025 says "we manually reviewed and labeled PII-related data", describes sampling production chat and "an AI data-generation pipeline", and links no training data.
- https://huggingface.co/api/datasets/Roblox/roblox-pii-classifier-benchmark recorded 2026-10-19
The dataset record describes "an evaluation-only dataset" that "accompanies Roblox PII Classifier v2.0", with a single test split of synthetic English conversations.
- https://huggingface.co/api/models/Roblox/roblox-pii-classifier-v2 recorded 2026-10-19
The Hugging Face API record reports "gated":false and the tag "license:apache-2.0", and lists "model.safetensors", "onnx/model.onnx", "inference.py" and "requirements.txt" among its files, with no training script.
- https://huggingface.co/Roblox/roblox-pii-classifier-v2/raw/main/README.md recorded 2026-10-19
The v2 card front matter declares "license: apache-2.0" and "base_model: FacebookAI/xlm-roberta-large". Its usage section runs the model with "python inference.py" after "pip install -r requirements.txt"; it mentions no training script, and says only that v2 "Broadened synthetic data generation and fine-tuning".
- https://huggingface.co/Roblox/roblox-pii-classifier/raw/main/README.md recorded 2026-10-19
The v1 card says "All the model training has been conducted with Roblox anonymized internal text datasets labeled by experts and AI generated conversations", and its front matter declares "license: apache-2.0".
Adoption
1 high confidenceAdoption is measured by Hugging Face downloads of the two releases Roblox publishes, v2 and the earlier v1. Roblox's own use of the model on its chat service is not counted, so the figure is a lower bound on use.
- https://huggingface.co/api/models/Roblox/roblox-pii-classifier recorded 2026-10-19
74 downloads in the trailing 30 days for Roblox/roblox-pii-classifier
- https://huggingface.co/api/models/Roblox/roblox-pii-classifier-v2 recorded 2026-10-19
244 downloads in the trailing 30 days for Roblox/roblox-pii-classifier-v2
Capability
3 medium confidenceOn privacy it outscores Llama Guard by a wide margin on every benchmark in Roblox's tables, and it reads the conversation around a message rather than the message alone. Privacy is the only harm it checks, while Llama Guard screens prompts and responses across fourteen hazard categories, so it sits a step below Llama Guard.
- https://about.roblox.com/newsroom/2025/11/open-sourcing-roblox-pii-classifier-ai-pii-detection-chat recorded 2026-10-19
The announcement says "PII Classifier receives a peak of over 200,000 queries per second on Roblox" with "<100ms P90 latency", served from an encoder-only model fine-tuned from XLM-RoBERTa-Large.
- https://huggingface.co/Roblox/roblox-pii-classifier-v2/raw/main/README.md recorded 2026-10-19
The v2 card lists three output labels, "privacy_asking_for_pii", "privacy_giving_pii" and "directing_users_off_platform", formats input as a target speaker "t" among other speakers, and reports expanding language coverage from 17 to 189 languages. Its F1 table gives 88.82% and 99.22% for v2 against 56.24% and 46.07% for LlamaGuard v3 8B on roblox-pii-safety-for-chat and Nemotron-PII.
Verified 2026-10-19, except some scores, still awaiting re-confirmation