Based on the paper on GPT-3, repetition seems to be one of the main problems with text synthesis: [2005.14165] Language Models are Few-Shot Learners (page 33)
However, this seems to be a habit in human chat in general. What I’ve learned is to train it a little better. Don’t use real world data. Use some kind of ideal conversation. Perhaps search for a movie script and conversation; I’d recommend David Mamet.
In the prompt, you may want to avoid prompts like “Eve is a friendly conversational bot” or “Eve is a regular teenager”. Try something wacky like “Eve is a salesperson who carries on conversations like her bonus depends on it.”