This model is part of the OPUS-MT project, an effort to make neural machine translation models widely available and accessible for many languages in the world. All models are originally trained using the amazing framework of Marian NMT, an efficient NMT implementation written in pure C++. The models have been converted to pyTorch using the transformers library by huggingface. Training data is taken from OPUS and training pipelines use the procedures of OPUS-MT-train.

Publications: OPUS-MT – Building open translation services for the World and The Tatoeba Translation Challenge – Realistic Data Sets for Low Resource and Multilingual MT (Please, cite if you use this model.)

@inproceedings{tiedemann-thottingal-2020-opus,
    title = "{OPUS}-{MT} {--} Building open translation services for the World",
    author = {Tiedemann, J{\"o}rg  and Thottingal, Santhosh},
    booktitle = "Proceedings of the 22nd Annual Conference of the European Association for Machine Translation",
    month = nov,
    year = "2020",
    address = "Lisboa, Portugal",
    publisher = "European Association for Machine Translation",
    url = "https://aclanthology.org/2020.eamt-1.61",
    pages = "479--480",
}

@inproceedings{tiedemann-2020-tatoeba,
    title = "The Tatoeba Translation Challenge {--} Realistic Data Sets for Low Resource and Multilingual {MT}",
    author = {Tiedemann, J{\"o}rg},
    booktitle = "Proceedings of the Fifth Conference on Machine Translation",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.wmt-1.139",
    pages = "1174--1182",
}

Model info

Release: 2021-02-23
source language(s): afr deu eng fry gos hrx ltz nds nld pdc yid
target language(s): afr deu eng fry nds nld
valid target language labels: >>afr<< >>ang_Latn<< >>deu<< >>eng<< >>fry<< >>ltz<< >>nds<< >>nld<< >>sco<< >>yid<<
model: transformer (base)
data: opus (source)
tokenization: SentencePiece (spm32k,spm32k)
original model: opus-2021-02-23.zip
more information released models: OPUS-MT gmw-gmw README
more information about the model: MarianMT

This is a multilingual translation model with multiple target languages. A sentence initial language token is required in the form of >>id<< (id = valid target language ID), e.g. >>afr<<

Usage

A short example code:

python

from transformers import MarianMTModel, MarianTokenizer

src_text = [
    ">>nld<< You need help.",
    ">>afr<< I love your son."
]

model_name = "pytorch-models/opus-mt-tc-base-gmw-gmw"
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)
translated = model.generate(**tokenizer(src_text, return_tensors="pt", padding=True))

for t in translated:
    print( tokenizer.decode(t, skip_special_tokens=True) )

# expected output:
#     Je hebt hulp nodig.
#     Ek is lief vir jou seun.

You can also use OPUS-MT models with the transformers pipelines, for example:

python

from transformers import pipeline
pipe = pipeline("translation", model="Helsinki-NLP/opus-mt-tc-base-gmw-gmw")
print(pipe(>>nld<< You need help.))

# expected output: Je hebt hulp nodig.

Benchmarks

test set translations: opus-2021-02-23.test.txt
test set scores: opus-2021-02-23.eval.txt
benchmark results: benchmark_results.txt
benchmark output: benchmark_translations.zip

langpair	testset	chr-F	BLEU	#sent	#words
afr-deu	tatoeba-test-v2021-08-07	0.674	48.1	1583	9105
afr-eng	tatoeba-test-v2021-08-07	0.728	58.8	1374	9622
afr-nld	tatoeba-test-v2021-08-07	0.711	54.5	1056	6710
deu-afr	tatoeba-test-v2021-08-07	0.696	52.4	1583	9507
deu-eng	tatoeba-test-v2021-08-07	0.609	42.1	17565	149462
deu-nds	tatoeba-test-v2021-08-07	0.442	18.6	9999	76137
deu-nld	tatoeba-test-v2021-08-07	0.672	48.7	10218	75235
eng-afr	tatoeba-test-v2021-08-07	0.735	56.5	1374	10317
eng-deu	tatoeba-test-v2021-08-07	0.580	35.9	17565	151568
eng-nds	tatoeba-test-v2021-08-07	0.412	16.6	2500	18264
eng-nld	tatoeba-test-v2021-08-07	0.663	48.3	12696	91796
fry-eng	tatoeba-test-v2021-08-07	0.500	32.5	220	1573
fry-nld	tatoeba-test-v2021-08-07	0.633	43.1	260	1854
gos-nld	tatoeba-test-v2021-08-07	0.405	15.6	1852	9903
hrx-deu	tatoeba-test-v2021-08-07	0.484	24.7	471	2805
hrx-eng	tatoeba-test-v2021-08-07	0.362	20.4	221	1235
ltz-deu	tatoeba-test-v2021-08-07	0.556	37.2	347	2208
ltz-eng	tatoeba-test-v2021-08-07	0.485	32.4	293	1840
ltz-nld	tatoeba-test-v2021-08-07	0.534	39.3	292	1685
nds-deu	tatoeba-test-v2021-08-07	0.572	34.5	9999	74564
nds-eng	tatoeba-test-v2021-08-07	0.493	29.9	2500	17589
nds-nld	tatoeba-test-v2021-08-07	0.621	42.3	1657	11490
nld-afr	tatoeba-test-v2021-08-07	0.755	58.8	1056	6823
nld-deu	tatoeba-test-v2021-08-07	0.686	50.4	10218	74131
nld-eng	tatoeba-test-v2021-08-07	0.690	53.1	12696	89978
nld-fry	tatoeba-test-v2021-08-07	0.478	25.1	260	1857
nld-nds	tatoeba-test-v2021-08-07	0.462	21.4	1657	11711
afr-deu	flores101-devtest	0.524	21.6	1012	25094
afr-eng	flores101-devtest	0.693	46.8	1012	24721
afr-nld	flores101-devtest	0.509	18.4	1012	25467
deu-afr	flores101-devtest	0.534	21.4	1012	25740
deu-eng	flores101-devtest	0.616	33.8	1012	24721
deu-nld	flores101-devtest	0.516	19.2	1012	25467
eng-afr	flores101-devtest	0.628	33.8	1012	25740
eng-deu	flores101-devtest	0.581	29.1	1012	25094
eng-nld	flores101-devtest	0.533	21.0	1012	25467
ltz-afr	flores101-devtest	0.430	12.9	1012	25740
ltz-deu	flores101-devtest	0.482	17.1	1012	25094
ltz-eng	flores101-devtest	0.468	18.8	1012	24721
ltz-nld	flores101-devtest	0.409	10.7	1012	25467
nld-afr	flores101-devtest	0.494	16.8	1012	25740
nld-deu	flores101-devtest	0.501	17.9	1012	25094
nld-eng	flores101-devtest	0.551	25.6	1012	24721
deu-eng	multi30k_test_2016_flickr	0.546	32.2	1000	12955
eng-deu	multi30k_test_2016_flickr	0.582	28.8	1000	12106
deu-eng	multi30k_test_2017_flickr	0.561	32.7	1000	11374
eng-deu	multi30k_test_2017_flickr	0.573	27.6	1000	10755
deu-eng	multi30k_test_2017_mscoco	0.499	25.5	461	5231
eng-deu	multi30k_test_2017_mscoco	0.514	22.0	461	5158
deu-eng	multi30k_test_2018_flickr	0.535	30.0	1071	14689
eng-deu	multi30k_test_2018_flickr	0.547	25.3	1071	13703
deu-eng	newssyscomb2009	0.527	25.4	502	11818
eng-deu	newssyscomb2009	0.504	19.3	502	11271
deu-eng	news-test2008	0.518	23.8	2051	49380
eng-deu	news-test2008	0.492	19.3	2051	47447
deu-eng	newstest2009	0.516	23.4	2525	65399
eng-deu	newstest2009	0.498	18.8	2525	62816
deu-eng	newstest2010	0.546	25.8	2489	61711
eng-deu	newstest2010	0.508	20.7	2489	61503
deu-eng	newstest2011	0.524	23.7	3003	74681
eng-deu	newstest2011	0.493	19.2	3003	72981
deu-eng	newstest2012	0.532	24.8	3003	72812
eng-deu	newstest2012	0.493	19.5	3003	72886
deu-eng	newstest2013	0.548	27.7	3000	64505
eng-deu	newstest2013	0.517	22.5	3000	63737
deu-eng	newstest2014-deen	0.548	27.3	3003	67337
eng-deu	newstest2014-deen	0.532	22.0	3003	62688
deu-eng	newstest2015-deen	0.553	28.6	2169	46443
eng-deu	newstest2015-ende	0.544	25.7	2169	44260
deu-eng	newstest2016-deen	0.596	33.3	2999	64119
eng-deu	newstest2016-ende	0.580	30.0	2999	62669
deu-eng	newstest2017-deen	0.561	29.5	3004	64399
eng-deu	newstest2017-ende	0.535	24.1	3004	61287
deu-eng	newstest2018-deen	0.610	36.1	2998	67012
eng-deu	newstest2018-ende	0.613	35.4	2998	64276
deu-eng	newstest2019-deen	0.582	32.3	2000	39227
eng-deu	newstest2019-ende	0.583	31.2	1997	48746
deu-eng	newstest2020-deen	0.604	32.0	785	38220
eng-deu	newstest2020-ende	0.542	23.9	1418	52383
deu-eng	newstestB2020-deen	0.598	31.2	785	37696
eng-deu	newstestB2020-ende	0.532	23.3	1418	53092

Acknowledgements

The work is supported by the European Language Grid as pilot project 2866, by the FoTran project, funded by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 771113), and the MeMAD project, funded by the European Union’s Horizon 2020 Research and Innovation Programme under grant agreement No 780069. We are also grateful for the generous computational resources and IT infrastructure provided by CSC -- IT Center for Science, Finland.

Model conversion info

transformers version: 4.12.3
OPUS-MT git hash: e56a06b
port time: Sun Feb 13 14:42:10 EET 2022
port machine: LM0-400-22516.local