What kinds of data are used to train your models, and what is the purpose of each type of data?

We use a combination of the following data sources:

• Publicly available and open-source datasets, primarily for research and development of new capabilities.
• Internally curated and annotated datasets created by trained specialists to improve model quality and reliability.
• User data, including information about product feature usage and quality signals, which is used to teach AI what types of outputs are liked most by users.
• Synthetically generated data, including model-generated examples and controlled system interaction data, which is used to improve specific capabilities for which annotated datasets are expensive or difficult to create.