WMT
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
WMT22 Ru-Sah news
58.2COMET
9
WMT22 En-Hr (news)
87.8COMET
9
WMT22 Uk-Cs news
90COMET
9
WMT22 Cs-Uk (news)
91.2COMET
9
WMT22 En-Ja (news)
89.5COMET
9
WMT22 En-De news
87COMET
9
WMT 2018 (newstest)
49.6BLEU
9
WMT En-Ro '16
39.9BLEU
9
WMT 2016 (test)
40.2BLEU
9
WMT en-de 2016 (newstest)
28.3BLEU
9
WMT17 Chinese-English (META-1)
69Ave %
9
WMT newstest 2014 (test)
15.24BLEU
9
WMT En-Fr (newstest2014)
37.5BLEU
9
WMT23 En-De (test)
29.57MCC
8
WMT23 Zh-En (test)
32.4MCC
8
WMT 19
0.509PRR
8
WMT14 T=0 (test)
23.1BLEU
8
WMT Ref-free European Languages
0.444Spearman Correlation
8
WMT Ref-free All Languages
44.5Spearman Correlation
8
WMT Ref-based European Languages
48.5Spearman Correlation
8
WMT Ref-based All Languages
47.7Spearman Correlation
8
WMT25 English → Italian (test)
64Group-by-Item Accuracy (tie calibration)
8
WMT English → Italian 2025 (test)
88Soft Pairwise Accuracy
8
WMT25 English-Ukrainian (official human annotations)
4.58Precision
8
WMT25 English-Italian (official human annotations)
10.2Precision
8