A system generates synthetic user profiles. The system receives various user profile parameters including a trajectory representing variation of relevance scores of epochs. The trajectory may be specified using any representation of a graph, including an image representation of a line graph, a set of tuples representing coordinates of the trajectory, or a natural language description of the trajectory. The user profile parameters may specify characteristics of epochs including the number of epochs, the aggregate time period of the epochs, and lengths of various epochs. The system generates a synthetic user profile using a machine learning-based language model. The system generates a user profile comprising a sequence of epochs such that the relevance scores of the sequence of epochs varies over time according to the specified trajectory. The synthetic user profiles may be used for training or evaluation of a machine learning-based models or systems based on such models.