Domingo mentioned looking at miplib2017 problems, which reminded me of this: I noticed some time ago that the current script offered with miplib2017 suffers from accuracy loss when checking the results, because it was parsing the solution file in printable form, which rounds values a bit. An example miplib result file is in
http://plato.asu.edu/ftp/milp_res1/benchmark.glpk.1threads.7200s.res where in the last column many solutions are reported as "fail" or "error".
I made an updated version that also uses a solution file in printable form, that offers more digits of accuracy (attached) and seems to work a lot better with a subset of miplib problems I used for testing.
Best Regards,
Chris Matrakidis
run_glpk.sh