Recently, LLMs pervaded the world of text understanding. Their ability to generate text based on instructions comes with the price of having an extremely large number of parameters and, consequently extremely large hardware demands or inference times. We consider that certain classification tasks, such as the detection of fixed lists of events from text can be achieved with acceptable performance even with smaller language models. The longstanding issue of collecting and annotating text is significantly eased by the presence of LLMs. Therefore, we employ a prompting technique to create a synthetic dataset of tweets annotated with events from the speaker’s life and train a small language model on it.