opus-2021-02-22.yml

release: deu-eng/opus-2021-02-22.zip
release-date: 2021-02-22
dataset-name: opus
modeltype: transformer-align
pre-processing: normalization + SentencePiece (spm32k,spm32k)
subwords:
   source: spm32k
   target: spm32k
subword-models:
   source: source.spm
   target: target.spm
source-languages:
   - deu
target-languages:
   - eng
training-data:
   deu-eng: Tatoeba-train (86845165) 
validation-data:
   deu-eng: Tatoeba-dev, 284809
   total-size-shuffled: 284809
   devset-selected: top 5000  lines of Tatoeba-dev.src.shuffled!
test-data:
   newssyscomb2009.deu-eng: 502/11821
   news-test2008.deu-eng: 2051/49380
   newstest2009.deu-eng: 2525/65402
   newstest2010.deu-eng: 2489/61724
   newstest2011.deu-eng: 3003/74681
   newstest2012.deu-eng: 3003/72812
   newstest2013.deu-eng: 3000/64505
   newstest2014-deen.deu-eng: 3003/67337
   newstest2015-ende.deu-eng: 2169/46443
   newstest2016-ende.deu-eng: 2999/64126
   newstest2017-ende.deu-eng: 3004/64399
   newstest2018-ende.deu-eng: 2998/67013
   newstest2019-deen.deu-eng: 2000/39282
   Tatoeba-test.deu-eng: 10000/81233
BLEU-scores:
   newssyscomb2009.deu-eng: 28.9
   news-test2008.deu-eng: 28.1
   newstest2009.deu-eng: 27.1
   newstest2010.deu-eng: 30.7
   newstest2011.deu-eng: 27.6
   newstest2012.deu-eng: 29.3
   newstest2013.deu-eng: 32.2
   newstest2014-deen.deu-eng: 34.3
   newstest2015-ende.deu-eng: 34.5
   newstest2016-ende.deu-eng: 40.6
   newstest2017-ende.deu-eng: 36.2
   newstest2018-ende.deu-eng: 44.3
   newstest2019-deen.deu-eng: 41.0
   Tatoeba-test.deu-eng: 53.5
chr-F-scores:
   newssyscomb2009.deu-eng: 0.555
   news-test2008.deu-eng: 0.551
   newstest2009.deu-eng: 0.546
   newstest2010.deu-eng: 0.588
   newstest2011.deu-eng: 0.557
   newstest2012.deu-eng: 0.571
   newstest2013.deu-eng: 0.584
   newstest2014-deen.deu-eng: 0.602
   newstest2015-ende.deu-eng: 0.603
   newstest2016-ende.deu-eng: 0.652
   newstest2017-ende.deu-eng: 0.614
   newstest2018-ende.deu-eng: 0.672
   newstest2019-deen.deu-eng: 0.649
   Tatoeba-test.deu-eng: 0.697