TechPrufer
Get started
Roblox
Roblox logoRoblox · Feature trail

Roblox Guard LLM Safety Toolkit

5 posts · Apr 2025 – May 2026

What this trail covers

Roblox Guard is an open-source safety toolkit for developers and platforms, featuring a state-of-the-art guardrail model for LLM safety and a dataset for safety benchmarking (Roblox Guard-Eval). This post details the evolution of Roblox's AI-powered content moderation systems over five years, emphasizing the integration of ML models, scalable infrastructure, and human experts to handle massive volumes of user-generated content (chat messages, voice communication, assets). The system prioritizes safety and civility, continuously growing and evolving with new technology. The voice safety classifier has been updated to increase parameters, expand to seven additional languages, and improve accuracy. The new model detects violations in Spanish, German, French, Portuguese, Italian, Korean, and Japanese, achieving a recall rate of 59.1% with low false-positive rates, and is optimized to serve up to 8,300 requests per second. The model architecture has been refined with an additional convolutional layer to reduce internal time resolution, resulting in a 20-30% faster inference time despite an increase in parameters. Training utilizes a large machine-labeled dataset of over 100,000 hours of speech, with data collection sampling abusive content at a higher probability to capture edge cases.

All Roblox trails
  1. CubePart: An Open-Vocabulary Part-Controllable 3D GeneratorIntroduced CubePart, a generative AI framework for open-vocabulary, part-controllable 3D mesh generation. CubePart outputs an assembled set of distinct, functional, and accurately-labeled meshes that match developer programming needs out of the box, expanding on the concept of fixed schemas. It allows creators to define a schema of required parts for an object, which then serves as an API contract for gameplay code, enabling interactive behavior without manual cleanup.
  2. State-of-the-Art LLM Helps Safeguard Unlimited Text Generation on RobloxAnnounces Roblox Guard 1.0, an open-source safety toolkit including a state-of-the-art guardrail model for LLM safety and the Roblox Guard-Eval dataset for safety benchmarking. This represents a new capability in safeguarding text generation.
  3. How Roblox Uses AI to Moderate Content on a Massive ScaleThis post details the evolution of Roblox's AI-powered content moderation systems over approximately five years, emphasizing the integration of ML models, scalable infrastructure, and human experts to handle massive volumes of user-generated content (chat messages, voice communication, assets). It highlights the system's principles: proactive moderation, real-time feedback, deploying AI only when it significantly outperforms humans in precision and recall at scale, and leveraging humans for continuous AI improvement, evolving and rare cases, complex investigations, and appeals. The system processes billions of creations and messages daily, with a median time to action of ten minutes for illegal content notices.
  4. Accelerating AI Inference for 3D Creation on RobloxRoblox implemented CUDA Graphs and KV caching to accelerate AI inference for 3D mesh generation, reducing token generation time from 60.5ms to 7.8ms and full object generation time from 31 seconds to 4 seconds. This optimization addresses latency bottlenecks identified during inference pipeline profiling, transforming the 3D generation experience from stop-and-wait to a more responsive and natural interaction.
  5. Launching More Languages for Our Open-Source Voice Safety ModelThe open-source voice safety classifier has been updated by increasing its parameters from 94.6 million to 120.2 million and expanding its language support to seven additional languages (Spanish, German, French, Portuguese, Italian, Korean, and Japanese). This has resulted in an improved recall rate of 59.1% on English-language voice chat data at a 1% false-positive rate, a 92% improvement over the previous release. The model is optimized to serve at scale, handling up to 8,300 requests per second. The architecture was refined with an additional convolutional layer to reduce internal time resolution, leading to a 20-30% faster inference time. Training now utilizes a machine-labeled dataset of over 100,000 hours of speech, with a focus on sampling abusive content.