In my local experiments. I've seen that with the training dataset that Falcon 7B is underperforming compared to Phi-2 or evan RAG (+ sementic similarity). But I'd like to know if somebody is doing goof with Falcon-7B. It seems that Falcon 7B is struggling with MCQ in the telecommunication area according to this paper https://arxiv.org/pdf/2402.15818. While playing with the 60 first training questions, Falcon is 30% accurate.