Yes, I just refreshed my memory by trying the mistake that I described earlier (just using data_x directly instead of the “shuffled” version that they create for you) and I get exactly the output you showed:
Iteration: 22000, Loss: 20.578871
Hutusaurus
Euca
Eustrioppn
Hocamptopanceus
Xuspeodon
Elacropechus
Uspeodon
last_name Uspeodon
They do the random shuffle just to make the results more interesting and less predictable. But of course things are still “predictable” here, because they set the random seeds exactly to make it easy to write test cases and the grader. You wouldn’t do that in a “real” system. 