At the French Ministry of Culture, we’ve built Compar:IA - a chatbot arena in the spirit of LMSYS, designed to benchmark and compare large language models in French through live user interactions. Over the past months, we’ve gathered 150,000+ lines of conversations and 100,000+ user preferences across models. This data is now publicly available on Hugging Face, and will soon be published on 
data.gouv.fr.
In this demo, we’ll open up the hood of this dataset and see how it could be used for post-training.
