> 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/playground.md).

# Test metric accuracy

Built-In Metrics and Custom Metrics can be tested for evaluation accuracy from the metric details screen.

The metrics playground can be opened from the metric details screen. Select `Evaluation Categories` from `Evaluation Library` in the sidebar, and open the details screen for the metric you want to test.

Use free input or AI-generated datasets to test whether the metric evaluates logs as expected.

### Select a test method

Metric evaluation accuracy can be tested using the following two methods.

<table><thead><tr><th width="181.5703125">Test method</th><th>Overview</th><th>Use case</th></tr></thead><tbody><tr><td>Free Input</td><td>Enter any input text or output text and check judgment details one by one</td><td>When you want to quickly check whether a metric can judge a specific case as intended</td></tr><tr><td>AI-generated dataset</td><td>Use multiple test data items generated by AI to check evaluation accuracy in bulk</td><td>When you want to check a metric's judgment tendencies and accuracy against various patterns</td></tr></tbody></table>

### Check evaluation accuracy with free input

With free input, you can enter any test data and check one by one whether the metric judges it as intended.

#### Set input content

For input-targeted metrics, enter the user prompt.

For output-targeted metrics, enter both the user prompt and the  response.

After filling out the details, click `Run Evaluation`.

#### Check evaluation results

When you run the evaluation, the evaluation result and latency for each model are displayed.

The evaluation results show whether the target metric judged the test data as Pass or Fail. For Remote metrics, you can compare the results across available models, such as the Fast Model and the High-Accuracy Model.

Latency shows the time required for evaluation by each model. Reviewing it together with the evaluation result lets you check the balance between evaluation accuracy and response time.

<figure><img src="/files/kjraNiuxmvtSsGF2WMGB" alt=""><figcaption></figcaption></figure>

### Check evaluation accuracy with an AI-generated dataset

With an AI-generated dataset, Citadel Radar automatically generates an evaluation dataset based on the specified use case and lets you check metric evaluation accuracy in bulk.

While manual input is ideal for verifying a single, specific case, AI-generated datasets are better suited for checking evaluation trends and accuracy across multiple test cases.

#### Create an evaluation dataset

Select the `AI-generated dataset` tab and create an evaluation dataset.

When creating an evaluation dataset, enter the target use case. For example, enter the usage scenario of the AI application you want to evaluate, such as a customer chatbot provided by a financial institution.

After entering the use case, select `Start Generation`.

<figure><img src="/files/g6zFYPF6vcWBHibgO7mP" alt="" width="375"><figcaption></figcaption></figure>

#### Check evaluation accuracy

When evaluation dataset generation is complete, the accuracy rate and evaluation result list are displayed.

The accuracy rate lets you check how correctly the metric judged the generated evaluation dataset.

<figure><img src="/files/BM0v7kLEUVse22OVYmcV" alt="" width="563"><figcaption></figcaption></figure>

#### Change the model and borderline handling

When checking the evaluation accuracy, you can adjust the model used and how borderline cases are handled.

By changing the model, you can compare evaluation trends across options like the High-Accuracy Model and the Fast Model.

For borderline handling, you can choose whether test data near the threshold is treated as Pass or Fail. Any changes to these settings are immediately reflected in the accuracy rate and the list of evaluation results.

#### Review specific cases

In the evaluation results list, you can review the user prompt, assistant response, ground truth label, metric judgment, and borderline status for each test case.

By comparing the ground truth labels with the metric judgments, you can identify which cases were evaluated correctly and where misclassifications occurred.

To view only misclassifications, enable `Show Misjudgements Only`.