opus4m-2020-08-12.yml

release: zle-eng/opus4m-2020-08-12.zip
release-date: 2020-08-12
dataset-name: opus4m
modeltype: transformer
pre-processing: normalization + SentencePiece (spm32k,spm32k)
subwords:
   - source: spm32k
   - target: spm32k
subword-models:
   - source: source.spm
   - target: target.spm
source-languages:
   - bel
   - orv
   - rue
   - rus
   - ukr
target-languages:
   - eng
test-data:
   - Tatoeba-test.bel-eng: 2500/15535
   - Tatoeba-test.multi-eng: 10000/58565
   - Tatoeba-test.orv-eng: 322/1713
   - Tatoeba-test.rue-eng: 120/566
   - Tatoeba-test.rus-eng: 10000/61769
   - Tatoeba-test.ukr-eng: 10000/55352
   - newstest2012.rus-eng: 3003/63778
   - newstest2013.rus-eng: 3000/56089
   - newstest2014-ruen.rus-eng: 3003/61219
   - newstest2015-enru.rus-eng: 2818/56934
   - newstest2016-enru.rus-eng: 2998/61076
   - newstest2017-enru.rus-eng: 3001/60714
   - newstest2018-enru.rus-eng: 3000/62925
   - newstest2019-ruen.rus-eng: 2000/37348
BLEU-scores:
   - Tatoeba-test.bel-eng: 38.7
   - Tatoeba-test.multi-eng: 50.7
   - Tatoeba-test.orv-eng: 7.7
   - Tatoeba-test.rue-eng: 13.1
   - Tatoeba-test.rus-eng: 53.4
   - Tatoeba-test.ukr-eng: 52.5
   - newstest2012.rus-eng: 32.0
   - newstest2013.rus-eng: 25.7
   - newstest2014-ruen.rus-eng: 28.8
   - newstest2015-enru.rus-eng: 27.5
   - newstest2016-enru.rus-eng: 26.8
   - newstest2017-enru.rus-eng: 30.3
   - newstest2018-enru.rus-eng: 26.3
   - newstest2019-ruen.rus-eng: 28.1
chr-F-scores:
   - Tatoeba-test.bel-eng: 0.573
   - Tatoeba-test.multi-eng: 0.660
   - Tatoeba-test.orv-eng: 0.227
   - Tatoeba-test.rue-eng: 0.335
   - Tatoeba-test.rus-eng: 0.682
   - Tatoeba-test.ukr-eng: 0.677
   - newstest2012.rus-eng: 0.585
   - newstest2013.rus-eng: 0.527
   - newstest2014-ruen.rus-eng: 0.570
   - newstest2015-enru.rus-eng: 0.547
   - newstest2016-enru.rus-eng: 0.543
   - newstest2017-enru.rus-eng: 0.568
   - newstest2018-enru.rus-eng: 0.542
   - newstest2019-ruen.rus-eng: 0.554