Turn insights into product excellence

Every user interaction is a chance to improve your product and the model’s ability to deliver better responses. You can export your logs as specific datasets (in CSV or JSONL format) for testing and offline evaluation. By identifying examples in your logs where quality and performance dipped (or excelled), you can build a reliable and reproducible baseline of expected results.

You can use these datasets for prompt refinement, performance tracking and more. For example, you can use the Gemini Batch API to run batch evaluations using datasets built up over time, see the Datasets Cookbook for an example. This allows you to test the changes to your Gemini model selection or application logic before you deploy them to users.

You also have the option to share specific datasets with Google to provide feedback on end-to-end model behavior for your specific use case. Shared datasets will be used to improve and develop Google products and services, including improving and training our models.