Aegis AI Content Safety LlamaGuard Permissive 1.0 EU AI Act Compliance Profile

Model Card

License

The use of this model is governed by the Llama 2 Community License Agreement.

Model Details

Llama Nemotron Safety Guard Permissive V1, formerly known as Aegis-AI-Content-Safety-LlamaGuard-LLM-Permissive-1.0, is an LLM content safety model. It is a parameter efficient instruction tuned version of Llama Guard based on Llama2-7B trained on Nvidia's content safety dataset Aegis Content Safety Dataset covering Nvidia's broad taxonomy of 13 critical safety risk categories.

Paper Details: Aegis 1.0: Online Adaptive AI Content Safety Moderation with Ensemble of LLM Experts

Model Description

The Llama-2-Nemotron-Safety-Guard-Permissive-7B-v1 model involves the following stages:

1. System instruction including the safety taxonomy, a safety policy with inclusions and, exclusions.
2. The system prompt instructs the LLM to moderate user prompt, partial dialog or full dialog.
3. The LLM response is a string which can be either safe or unsafe. If the string generated by the LLM is "unsafe", on a new line, the category ID of violation is output by the LLM based on the policy in the system prompt.
4. Novel safety risk categories and policy can be provided in the instruction for the model to categorize using the novel taxonomy and policy.
5. The safety taxonomy and policy used to train the models contain 13 critically unsafe risk categories, a safe category and a "needs caution" category.
6. Internally annotated dataset called Nemotron Content Safety Dataset V1 of approximately 11,000 prompts and responses are used to instruction tune the model. Annotations are at dialog level not per turn. We have since collected in tot

Your obligations as a Aegis AI Content Safety LlamaGuard Permissive 1.0 deployer

Track your progress

Sign up to track obligations, receive deadline reminders, and generate audit packages.

Tags

Languages

Library