Luciole RAG is a supervised fine-tuning dataset for retrieval-augmented generation, built to train the Luciole models. Each example is a chat conversation where the assistant answers a question using only a set of retrieved document chunks given in the system prompt, quotes and cites its sources, and declines to answer when the documents do not contain the answer.
It contains two subsets derived from existing question-answering benchmarks:
