Converted older .tsv from pdf-extraction comparison work with: cat 1k_random_identified_combined.tsv | jq -c --slurp --raw-input --raw-output 'split("\n") | .[:-1] | map(split("\t")) | map({"doi": .[0], "title": .[1], "authors": ( .[2] | split(";") ), "year": .[3], "journal": .[4], "publisher": .[5], "subject": .[6], "type": .[7], "sha": .[8]}) | .[]' > crossref_sample.bibjson Note that neither bibjson file is a superset of the either: 944 unique SHA1 which exist in both lists 964 in crossref_sample.sha1 979 in grobid_sample.sha1 So scoring should be on a basis of "out of 944 lines". If this is confusing we can trim the files down.