linguardia/opus-mt-en-nds

English → Low German (opus-mt finetune)

Translates English into Low German. Finetuned by Linguardia from Helsinki-NLP/opus-mt-en-mul on English↔Low German sentence pairs from Tatoeba.

Built to generate course material for a language that no major machine translator supports.

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("linguardia/opus-mt-en-nds")
model = AutoModelForSeq2SeqLM.from_pretrained("linguardia/opus-mt-en-nds")

batch = tok([">>nds<< I wake up early."], return_tensors="pt")
out = model.generate(**batch, num_beams=4, max_length=64)
print(tok.batch_decode(out, skip_special_tokens=True))

The target language is selected by the >>nds<< prefix on the source text. It was added as a new token rather than replacing one of the base model's existing languages, so every language the base already handled still works.

Training

base	Helsinki-NLP/opus-mt-en-mul
training pairs	5,157
held out	331 dev, 19 gold
epochs	5.0
batch size	32
optimiser	Adafactor, lr 5e-5, 500 warmup steps
max sequence	48 tokens (99.9% of source, 99.6% of target)
hardware	Apple M4, MPS

Evaluation

Held-out test set (1,200 pairs, never trained on, and filtered so no test sentence shares either its English or its Low German side with any training pair):

metric	score
BLEU	39.5
chrF	59.3

For comparison, the models already published from this pipeline score chrF 52.8 (Tachelhit), 41.4 (Scots) and 34.4 (Old English) on their own held-out sets.

Course built from it

Translating the canonical 2,000-sentence course corpus:

metric	nds	shi	kab	jbo	tok
coverage %	99.5	89.0	87.8	97.8	78.7
rules	1467	1257	1373	1318	1303
attributed %	50.4	28.3	26.4	34.3	18.7

Data

5,157 English-Low German pairs from Tatoeba (CC-BY 2.0 FR), being every linked pair minus those overlapping the held-out test set. North Low Saxon in Sass orthography, which is what the Tatoeba nds corpus is.

No synthetic or model-generated training data was used. Base model Helsinki-NLP/opus-mt-en-mul, Apache-2.0.

What it is not

Machine output that no Low German speaker has read. Every row ships verified: false; corrections from speakers are what flip it.

Training data and attribution

Sentence pairs come from Tatoeba, released under CC-BY 2.0 FR, with some sentences under CC0 1.0. Tatoeba's sentences are written by volunteers, and this model is a derivative of their work: if you use it, credit Tatoeba and its contributors.

Pairs were deduplicated on exact (English, Low German) text. No machine translation was used to create the training data at any point.

Licence chain

layer	licence
base model Helsinki-NLP/opus-mt-en-mul	Apache-2.0
training data	CC-BY 2.0 FR / CC0 1.0
these weights	Apache-2.0, with the attribution above

Limitations

• Trained on short everyday sentences. Expect it to be weakest on long, technical or literary text.
• Neural MT translates each sentence independently, so a term may be rendered differently across a corpus. Check consistency if that matters to you.
• Output should be reviewed by a speaker before being taught to learners.