Relationships are declared once and mirrored automatically. A
shares records pair contains some of the same recordings, so
training on one and evaluating on the other contaminates the test set.
Verified means the overlap was checked against the actual data files;
otherwise it is taken from documentation.
The cpsc_2018 cohort holds 6,877 records, exactly the size of the CPSC-2018 public training set, renamed A0001-A6877; cpsc_2018_extra adds 3,453 Q#### records that CPSC-2018 did not release publicly. Assume the 6,877 overlap completely and do not evaluate on CPSC-2018 after training here. The count agreement is strong evidence but was not checked against CPSC-2018 files — unlike the other four cohort overlaps on this page.