Icebreaker Question Generation
Icebreaker questions are often hard to answer, but it is (arguably) even harder to come up with good ones. This app both generates a list of icebreakers given a type (e.g., creative, funny, deep), then judges the difficulty of those questions using a classifier. Another use case of this app could be to better understand potential users of a product. The users' answers could help a company with asking more targeted survey questions and the like.
The question generation uses LM prompting on the `xlarge` Cohere pretrained LM. The difficulty classifier is an SVM that inputs embeddings generated through Cohere's API. The code for querying the model APIs and serving the web app is only 135 lines long—showing that anyone could build it, irrespective of coding ability.