This PR addresses #1265 (closed) . I have found the following inconsistencies between the output of buildmaster and the files stored in nnpdfcpp.
A) DATA files: no inconsistencies, except for ZpT data (see below) and for the fact that a slightly different formatting is used sometimes (possibly due to the generation of the DATA files on different machines). -> No action taken.
B) SYSTYPE files: I have found inconsistencies for the following experiments:
NMC: the four last uncertainties should be MULT, as in the SYSTYPE file stored in nnpdfcpp -> corrected in buildmaster;
ATLAS_DY_2D_8TEV_LOWMASS: there is no luminosity uncertainty (the cross section is normalised); the SYSTYPE file in nnpdfcpp is correct -> corrected buildmaster;
ATLAS_WP_JET_8TEV_PT and ATLAS_WM_JET_8TEV_PT: the non-perturbative uncertainty should be SKIP, as non-perturbative corrections are not applied; the SYSTYPE file in nnpdfcpp is correct -> corrected buildmaster;
ATLAS_TTB_DIFF_8TEV_LJ_TPTNORM, ATLAS_TTB_DIFF_8TEV_LJ_TRAPNORM, ATLAS_TTB_DIFF_8TEV_LJ_TTRAPNORM and ATLAS_TTB_DIFF_8TEV_LJ_TTMNORM: an uncertainty was incorrectly assumed as MULT while it should be ADD; the implementation in buildmaster is correct -> corrected files stored in nnpdfcpp.
ATLASZPT7TEV, ATLASZPT8TEVMDIST, ATLASZPT8TEVYDIST, CMSZDIFF12: an extra theoretical uncertainty is appended in the DATA and SYSTYPE files stored in nnpdfcpp. These files are generated from the buildmaster output by running the script process_zpt.sh in nnpdf/nnpdfcpp/data/commondata/31systematics. Therefore the buildmaster and nnpdfcpp files ought to be different. This way of processing the files for these datasets is a remnant of NNPDF3.1. -> No action taken (except to check that the script reproduces the files stored in nnpdfcpp from the output of buildmaster).
Note that I have checked only the datasets that we used or that are going to be used for NNPDF4.0. There might be inconsistencies in deprecated/outdated files. In this respect, I think that the best thing to do is to remove them, a separate issue on a general cleanup is perhaps advisable before making the code public.
Note also that I've realised that the ATLASPHT15 and ATLASPHT12 datasets specified in the template runcards and used so far should be replaced by the variants ATLASPHT15_SF and ATLASPHT12_SF, which are the only ones produced by buildmaster.
In summary, all the inconsistencies had to be fixed at the level of buildmaster, except for the ATLAS photon data and for the ATLAS 8 TeV top-pair differential data. These changes seem however inconsequential if one compares the chi2 obtained after these changes (see https://vp.nnpdf.science/diPHOGXBToSdC8K02654sg==) with the chi2 obtained before these changes (see https://vp.nnpdf.science/NQzde1SETUuB-Ma677wMrg==). I therefore conclude that the weighted fits do not need to be repeated.
Finally, note that his PR also includes an update of the template runcards, in which the correct denomination of the surviving ATLASPHT15_SF dataset has been updated and the datasets that we decided to discard from the baseline fit (as an outcome of the recent weighted fit campaign) have been removed.
