When were the training data collected?

Data collection is ongoing, as we make regular updates to AI models. Details vary based on the language and capability being modeled, but we typically use training data that was collected up to 24 months before the model training begins. However, we use our small, internally annotated datasets for longer than 24 months because they have been manually verified to be free of personal information and the data is costlier to produce than the other types of data we use.