> For the complete documentation index, see [llms.txt](https://docs.radar.citadel-ai.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.radar.citadel-ai.com/en/application-owner-guide/evaluation-library/built-in-metrics.md). # Review Built-In Metrics Built-in metrics are metrics designed by Citadel AI with reference to external AI safety guidelines and frameworks, and are included in Citadel Radar in advance. Built-in metrics are organized by Evaluation Category. When you select an Evaluation Category, you can review the Evaluation Metrics included in that category and related AI safety framework items. {% hint style="info" %} Supported AI safety frameworks will be expanded over time. If there is an AI safety framework you would like supported, please contact . {% endhint %} ### What you can check with built-in metrics With built-in metrics, you can evaluate LLM application inputs and outputs based on criteria such as the following. * Whether they encourage bias or discrimination against people with specific attributes * Whether they include information that should not be disclosed outside of the organization * Whether users are entering inappropriate or dangerous instructions to the LLM By using built-in metrics, you can start evaluation based on common risk domains without creating your own evaluation criteria. {% hint style="info" %} Before reviewing built-in metrics, confirm the relationship between Evaluation Categories and Evaluation Metrics. For details, see [Manage the Evaluation Library](/en/application-owner-guide/evaluation-library.md). {% endhint %} ### Review the list of Evaluation Categories Select `Evaluation Categories` in the `Evaluation Library` section of the sidebar. In the Evaluation Category list, you can review each category's name, target, description, number of related frameworks, and number of included metrics.

If you select, for example, `Disclosure of Personal Information` from the Evaluation Categories, you can review the Evaluation Metrics included in that category and the linked AI safety framework items. #### Review Evaluation Metrics included in an Evaluation Category

The `Disclosure of Personal Information` category is defined as "*The risk of personally identifiable information (PII) being disclosed to or inferred by third parties without consent through LLM responses.*" In other words, this category handles risks caused by outputs rather than the input itself. Therefore, the following metrics are defined as built-in metrics in this category.

Target	Overview
Input	Evaluates whether the input attempts to elicit information from the LLM for the purpose of obtaining, inferring, or disclosing a third party's personal information
Output	Evaluates whether the generated output actually contain third-party personal information.

{% hint style="info" %} Since the `Disclosure of personal information` category addresses risks arising from the outputs, it does not check if the input itself contains personal information. Instead, it evaluates whether the input intentionally attempts to elicit such outputs. For details about this relationship between categories and metrics, see [Manage the Evaluation Library](/en/application-owner-guide/evaluation-library.md). {% endhint %} For some metrics, you can select `Local` or `Remote` as the model type. `Local` uses an SLM running in an environment hosted by Citadel AI. `Remote` uses a Remote LLM service provided by a third party. When you select each metric, you move to the playground screen. For more information on how to use the playground, see [Test metric accuracy](/en/application-owner-guide/evaluation-library/playground.md). #### Review AI safety framework items linked to an Evaluation Category On the Evaluation Category details screen, you can also review AI safety framework items related to that category. When you select each framework item, you can review the item overview and related Evaluation Metrics.