This is a demo of Baskerville on lmsys/chatbot_arena_conversations, a dataset of real-world conversations where a user asks 2 LLMs a question and votes on which they prefer. Baskerville uses interpretability to find these patterns in which responses users prefer/disprefer, the way a data scientist or product analyst would. This can be applied to any dataset. See this post for more context. Request a demo here.
Take feature 2233, for example. It represents the concept of asking "let me know if you need anything more", appears 688 times in the dataset, and is in the winning response in 464 of those samples. This strong correlation suggests some pattern in user preferences.
This tool shows only "featured features" by default. You can toggle this off to explore more on your own.