This is a reminder to update the RSG readme when I'm at a computer. Currently it notes that the data is one full trajectory split into chunks, but it doesn't mention how those chunks were split into train/valid/test. After checking with some folks, the answer is randomly since there are significant distribution shifts over time, making this an interpolation-in-time task rather than extrapolation.