jkkma/opus-mt-tc-bible-big-mul-deu_eng_nld-onnx

opus-mt-tc-bible-big-mul-deu_eng_nld, exported to ONNX

An ONNX export of Helsinki-NLP/opus-mt-tc-bible-big-mul-deu_eng_nld at revision bb1ef830d540449c89c7ee5b9ea5b1fc666db3d5, in the merged-decoder layout, float32. Produced for Uindosill and published so that the exact bytes its measurements describe can be downloaded and checked, rather than only reproduced.

Modifications, per Apache-2.0 section 4(b)

The original Marian checkpoint was exported to ONNX by scripts/export-translation-onnx.py in the Uindosill repository, which splits it into an encoder graph and a decoder graph with past key values exposed behind a use_cache_branch switch. The weights are unchanged and unquantised — float32 in, float32 out. What changed is the container and the graph split. This repository redistributes the exported graphs; it does not redistribute the original checkpoint.

A copy of the Apache License 2.0 ships beside these files as LICENSE, per section 4(a).

Attribution notices retained from the source, per Apache-2.0 section 4(c)

The upstream repository was read at revision bb1ef830d5 on 2026-08-20 — its file listing and every text file in it. It ships no NOTICE file, so section 4(d) has nothing to reproduce, and it carries no copyright, patent or trademark notice anywhere, so none is reproduced here rather than one being invented. That negative result is stated rather than left silent, because a card that omits a NOTICE file and one that records there is none read identically to anyone downstream.

The attribution notices it does carry are these:

• Developed by the Language Technology Research Group at the University of Helsinki, as part of the OPUS-MT project. Originally trained with Marian NMT and converted to PyTorch with the transformers library; training data from OPUS, training pipelines from OPUS-MT-train.
• Original model: opusTCv20230926max50+bt+jhubc_transformer-big_2024-08-18.zip
• The source asks to be cited. If you use this export, cite the base model's publications:

@article{tiedemann2023democratizing,
  title={Democratizing neural machine translation with {OPUS-MT}},
  author={Tiedemann, J{\"o}rg and Aulamo, Mikko and Bakshandaeva, Daria and Boggia, Michele and Gr{\"o}nroos, Stig-Arne and Nieminen, Tommi and Raganato, Alessandro and Scherrer, Yves and Vazquez, Raul and Virpioja, Sami},
  journal={Language Resources and Evaluation},
  number={58},
  pages={713--755},
  year={2023},
  publisher={Springer Nature},
  issn={1574-0218},
  doi={10.1007/s10579-023-09704-w}
}

@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",
}

Acknowledgements

Reproduced from the base model's card:

The work is supported by the HPLT project, funded by the European Union’s Horizon Europe research and innovation programme under grant agreement No 101070350. We are also grateful for the generous computational resources and IT infrastructure provided by CSC -- IT Center for Science, Finland, and the EuroHPC supercomputer LUMI.

What is here

Nine files, 1,435,604,524 bytes.

file	bytes
encoder_model.onnx	545,922,847
decoder_model_merged.onnx	886,620,027
config.json	1,056
generation_config.json	308
source.spm	736,809
target.spm	808,244
vocab.json	1,514,254
tokenizer_config.json	900
special_tokens_map.json	79

decoder_model_merged.onnx does not export deterministically. Exported twice on one machine, minutes apart, same interpreter and same cached checkpoint, it comes out the same size and a different SHA-256. Eight of the nine files reproduce exactly; that one does not. So these bytes are an artefact that was built once and published, rather than one anybody can rebuild byte-identically from the script — which is the reason this repository exists at all.

Three things that are easy to get wrong

• >>eng<< is mandatory, and it is one token — vocabulary entry 693, not a punctuation sequence to be segmented. Every source string must begin with it. Given Spanish without it, this checkpoint returns fluent German rather than an error, so a forgotten prefix produces confident output in the wrong language that nothing downstream will catch.
• decoder_start_token_id is 58433, which is also pad_token_id, and it is in bad_words_ids. Three roles for one id.
• The tokenizer declares model_max_length 512 while config.json says max_position_embeddings 1024. 512 is the smaller and is the number to design against.

generation_config.json says num_beams: 4. Every figure below was produced at beam 6, passed explicitly, batch 1.

Measured quality, into English

FLEURS test in full, 8,149 sentences across 24 languages, beam-6, batch 1, on one desktop CPU, 2026-08-20. chrF++ at nrefs:1|case:mixed|eff:yes|nc:6|nw:2|space:no|version:2.6.0.

The source-copy floor is what a hypothesis scores by echoing its own untranslated source. It is per language because it is a property of the pair — a Cyrillic source shares almost no character n-grams with an English reference, a French one shares many — so a single bar across all of them would be a different test in each.

	chrF++	source-copy floor	margin	required
pt	68.52	21.77	+46.75	+23.23	pass
da	67.66	21.64	+46.02	+23.36	pass
sv	66.43	20.91	+45.52	+24.09	pass
mt	65.22	19.42	+45.8	+25.58	pass
ro	64.52	21.76	+42.76	+23.24	pass
fr	63.95	23.1	+40.85	+21.9	pass
de	63.64	20.78	+42.86	+24.22	pass
bg	62.66	2.13	+60.53	+42.87	pass
cs	60.88	15.96	+44.92	+29.04	pass
uk	59.42	2.0	+57.42	+43.0	pass
hr	59.0	16.44	+42.56	+28.56	pass
et	58.58	16.84	+41.74	+28.16	pass
it	58.06	22.42	+35.64	+22.58	pass
lv	57.67	14.54	+43.13	+30.46	pass
ru	57.61	2.1	+55.51	+42.9	pass
nl	57.53	22.07	+35.46	+22.93	pass
hu	56.75	15.45	+41.3	+29.55	pass
fi	56.67	15.92	+40.75	+29.08	pass
sl	56.23	16.42	+39.81	+28.58	pass
es	56.17	21.4	+34.77	+23.6	pass
el	54.42	2.37	+52.05	+42.63	pass
lt	54.35	15.47	+38.88	+29.53	pass
pl	53.47	15.54	+37.93	+29.46	pass
sk	44.26	16.11	+28.15	+28.89	FAIL

Slovak fails, by 0.74. It is the one language the publishing project's own bar rejects, and it was the outlier its research predicted, being absent from the sibling opus-mt-mul-en card's source list.

Zero degenerate collapses in 8,149 sentences. 31 trailing-punctuation runs — hypotheses that finish translating correctly and then will not stop emitting . . . . — concentrated in Maltese (17) and Greek (5).

What these numbers do not cover

• They are not a cascade figure. FLEURS transcripts go in; no speech recognition runs. Real use puts ASR output through this model, and ASR output is not clean written prose. Any figure here is a lower bound on the penalty a real pipeline pays.
• Numbers spoken as words are a known weak point. German neunzehnhundertneunundzwanzig (1929, as a speaker would say it) came back as "the nineteenth century". Digits survive; compound number words do not.
• The language list is membership, not capability. The base model's own card disclaims it: "for a large number of language pairs it will not work at all."
• English is a passthrough and is not scored — it returns byte-identical input.
• No human has rated adequacy. The publishing project's quality gate has a second criterion, a human adequacy check, and it is unperformed. That gate is not passed.

Provenance

Exported with torch 2.13.0+cpu, transformers 4.57.6, optimum 2.1.0, onnx 1.22.0, onnxruntime 1.29.0, sentencepiece 0.2.2. Full detail, and everything above with its caveats, is in docs/UNPROVEN.md under Translating into English.