VAMOS navigation-only (without natural language preference annotations nor co-training VQA data) dataset with 100% of TartanDrive 2 data, 50% of SCAND data, 25% of CODa data, and 100% of in-domain spot data. Whenever daatsets aren't 100%, they are ranked by curvature and output of length 5.