from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

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

batch = tok([">>rhg<< 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 >>rhg<< 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` (77M params)
training pairs	3,548
held out	300 dev, 0 gold
epochs	20.0
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 Tatoeba (300 pairs never seen in training):

metric	score
BLEU	18.4
chrF	42.5

Translating the same 2,000-sentence course corpus:

corpus	rows	distinct	vocab	words/sentence
this model	2000	1998 (99.9%)	2246	6.7

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, Rohingya) 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 (Tatoeba)	CC-BY 2.0 FR / CC0 1.0
these weights	Apache-2.0, with the attribution above

Limitations

Trained on Tatoeba, which is 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.