Comment by pseudometa

When the story input parameters are too close or similar to an existing work, you'll receive output containing training artifacts. For example if the parameters for a story included topics such as a castle in Norway featuring a princess with powers. While the overall story will be its own thing, you'll definitely get names of characters and story elements from Disney's Frozen. The main ways to avoid this is to increase the top_p and temperature of the LLM requests, as well as add additional specifics to the story prompt. Are there any other techniques out there that would help in this?