Llama-3.1-8B-Instruct-RobloxGuard-1.0

<h1 align="center">Roblox Guard 1.0: Advancing Safety for LLMs with Robust Guardrails</h1> <div align="center" style="line-height: 1;"> <a href="https://huggingface.co/Roblox/Llama-3.1-8B-Instruct-RoGuard-1.0" target="_blank"><img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-RobloxGuard 1.0-ffc107?color=ffc107&logoColor=white"/></a> <a href="https://github.com/Roblox/RobloxGuard-1.0"><img alt="github" src="https://img.shields.io/badge/🤖%20Github-RobloxGuard%201.0-ff6b6b?color=1783ff&logoColor=white"/></a> <a href="https://github.com/Roblox/RobloxGuard-1.0/blob/main/LICENSE"><img src="https://img.shields.io/badge/Model%20License-RAIL_MS-green" alt="Model License"></a> </div> <div align="center" style="line-height: 1;"> <a href="https://huggingface.co/datasets/Roblox/RobloxGuard-Eval" target="_blank"><img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-RobloxGuardEval-ffc107?color=1783ff&logoColor=white"/></a> <a href="https://creativecommons.org/licenses/by-nc-sa/4.0/"><img src="https://img.shields.io/badge/Data%20License-CC_BY_NC_4.0-blue" alt="Data License"></a> </div> <div align="center" style="line-height: 1;"> <a href="https://corp.roblox.com/newsroom/2025/07/roblox-guard-advancing-safety-for-llms-with-robust-guardrails" target="_blank"><img src=https://img.shields.io/badge/Roblox-Blog-000000.svg?logo=Roblox height=22px></a> <a href="https://arxiv.org/abs/2512.05339" target="_blank"><img src="https://img.shields.io/badge/Paper-2512.05339-b5212f.svg?logo=arxiv" height="22px"></a> </div>

Roblox Guard 1.0, a SOTA instruction fine-tuned LLM, is designed to help safeguard our Text Generation API. It performs safety classification at both the prompt and response levels, deciding whether or not each input or output violates our policies. This dual-level assessment is essential for moderating both user queries and the model’s own generated outputs. At the heart of our system is an LLM that’s been fine-tuned from the Llama-3.1-8B-Instruct model, the license for which is at: https://www.llama.com/llama3_1/license/. We trained this LLM with a particular focus on high-quality instruction tuning to optimize for safety judgment performance.

📊 Model Benchmark Results

We benchmark Roblox Guard 1.0 model on a comprehensive set of open-source datasets for both prompt and response, as well as on RobloxGuard-Eval. This allows us to evaluate our model on both in-domain and out-of-domain datasets. We report our results in terms of F-1 score for binary violating/non-violating classification. In the table above, we compare our performance with that of several well-known models. The RobloxGuard 1.0 outperforms other models while generalizing on out-of-domain datasets.

• Prompt Metrics: These evaluate how well the model classifies or responds to potentially harmful user inputs
• Response Metrics: These measure how well the model handles or generates responses, ensuring its outputs are safe and aligned.

🔗 GitHub Repository

You can find the full source code and evaluation framework on GitHub:

👉 Roblox/RobloxGuard on GitHub

Citation

If you are using this model, please cite it as:

@article{nandwana2025taxonomy,
  title={Taxonomy-Adaptive Moderation Model with Robust Guardrails for Large Language Models},
  author={Nandwana, Mahesh Kumar and Lim, Youngwan and Liu, Joseph and Yang, Alex and Notibala, Varun and Khanna, Nishchaie},
  journal={arXiv preprint arXiv:2512.05339},
  year={2025}
}