I don't understand the model selection in checkpoint. The results published in the paper are based on Vicuna, but the checkpoint is based on Mistral?