Aegis AI Content Safety

weights

adapter-open(HF) but base LlamaGuard gated via Meta access

data

Aegis safety dataset documented

license

Llama-2-Community(NOT OSI)

NVIDIA publishes the adapter weights on Hugging Face without an access gate, under the Llama 2 Community License, which carries use restrictions and is not OSI-approved. The adapter runs on top of Meta's LlamaGuard-7b, whose download needs Meta's approval, and the Aegis training dataset is released separately under CC-BY-4.0.

• https://huggingface.co/api/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-1.0 recorded 2026-09-29

  The Hugging Face API record for the training dataset carries the tag "license:cc-by-4.0", lists two parquet files including "Content Moderation Extracted Annotations 02.08.24_train_release_0418_v1.parquet", and describes it as "an open-source content safety dataset (CC-BY-4.0)".

• https://huggingface.co/api/models/meta-llama/LlamaGuard-7b recorded 2026-09-29

  The Hugging Face API record for Meta's base model meta-llama/LlamaGuard-7b carries the tag "license:llama2" and marks the repository gated with the value manual, so the base weights the adapter loads require Meta's access approval; the record's tags include "text-generation-inference".

• https://huggingface.co/api/models/nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0 recorded 2026-09-29

  The Hugging Face API record for the Defensive adapter carries the tags "license:llama2" and "dataset:nvidia/Aegis-AI-Content-Safety-Dataset-1.0", sets gated to false, and lists the downloadable weights file "adapter_model.safetensors"; its PEFT config points the adapter at a base of "models--meta-llama--LlamaGuard-7b".

• https://huggingface.co/nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0 recorded 2026-09-29

  The model page's card reads "The use of this model is governed by the" Llama 2 Community License Agreement and describes the adapter as "a parameter efficient instruction tuned version of" Llama Guard, trained on NVIDIA's content safety dataset covering "13 critical safety risk categories".

• https://huggingface.co/nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0/raw/main/README.md recorded 2026-09-29

  The model card states "The use of this model is governed by the" Llama 2 Community License Agreement, describes the model as "a parameter efficient instruction tuned version of" Llama Guard based on Llama2-7B, and says the adapter was trained with PEFT and "the training and validation code from the" Llama recipes repository.

Adoption is measured by Hugging Face downloads for the one artifact NVIDIA has declared, nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0. No other channel is declared for the model, so this single download count is the only visible measure of its use.

• https://huggingface.co/api/models/nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0 recorded 2026-08-13

  4,303 downloads in the trailing 30 days for nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0

Aegis covers 13 risk categories and accepts custom categories in the prompt, but it was trained only to classify English prompts, labeling responses through the same template. Llama Guard 3 classifies prompts and responses across 14 categories in eight languages and also screens tool calls, so Aegis sits a step below it.

• https://huggingface.co/api/models/nvidia/Aegis-AI-Content-Safety-LlamaGuard-Permissive-1.0 recorded 2026-09-29

  The Hugging Face API record for the companion adapter "nvidia/Aegis-AI-Content-Safety-LlamaGuard-Permissive-1.0" carries the same tag "dataset:nvidia/Aegis-AI-Content-Safety-Dataset-1.0" and ships its own "adapter_model.safetensors", so NVIDIA still publishes a permissive variant beside the defensive one.

• https://huggingface.co/nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0 recorded 2026-09-29

  The model page's card reads "The use of this model is governed by the" Llama 2 Community License Agreement and describes the adapter as "a parameter efficient instruction tuned version of" Llama Guard, trained on NVIDIA's content safety dataset covering "13 critical safety risk categories".

• https://huggingface.co/nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0/raw/main/README.md recorded 2026-09-29

  The model card says the safety policy used for training contains "13 critically unsafe risk categories, a safe category and a" needs-caution category, and that "Novel safety risk categories and policy can be provided in the instruction", but warns "Model has only been trained to perform prompt classification since the annotations were not available at turn level"; the front matter lists only language en, and the evaluation tables compare it with "Llama Guard Base".