opus-2021-02-18.yml

release: fra-ces/opus-2021-02-18.zip
release-date: 2021-02-18
dataset-name: opus
modeltype: transformer
pre-processing: normalization + SentencePiece (spm32k,spm32k)
subwords:
   source: spm32k
   target: spm32k
subword-models:
   source: source.spm
   target: target.spm
source-languages:
   - fra
target-languages:
   - ces
training-data:
   ces-fra: Tatoeba-train (35981979) 
validation-data:
   ces-fra: Tatoeba-dev, 1000
   total-size-shuffled: 1000
   devset-selected: top 1000  lines of Tatoeba-dev.src.shuffled!
test-data:
   newssyscomb2009.fra-ces: 502/10032
   news-test2008.fra-ces: 2051/42484
   newstest2009.fra-ces: 2525/55533
   newstest2010.fra-ces: 2489/52958
   newstest2011.fra-ces: 3003/65653
   newstest2012.fra-ces: 3003/65456
   newstest2013.fra-ces: 3000/57250
   Tatoeba-test.fra-ces: 425/2625
BLEU-scores:
   newssyscomb2009.fra-ces: 22.0
   news-test2008.fra-ces: 20.1
   newstest2009.fra-ces: 20.6
   newstest2010.fra-ces: 20.4
   newstest2011.fra-ces: 21.8
   newstest2012.fra-ces: 20.6
   newstest2013.fra-ces: 22.0
   Tatoeba-test.fra-ces: 46.7
chr-F-scores:
   newssyscomb2009.fra-ces: 0.501
   news-test2008.fra-ces: 0.484
   newstest2009.fra-ces: 0.486
   newstest2010.fra-ces: 0.489
   newstest2011.fra-ces: 0.495
   newstest2012.fra-ces: 0.478
   newstest2013.fra-ces: 0.488
   Tatoeba-test.fra-ces: 0.672