Helsinki-NLP/opus-mt_tiny_deu-eng

OPUS-MT-tiny-deu-eng

Distilled model from a Tatoeba-MT Teacher: OPUS-MT-models/de-en/opus-2020-02-26, which has been trained on the Tatoeba dataset.

We used the OpusDistillery to train new a new student with the tiny architecture, with a regular transformer decoder. For training data, we used Tatoeba. The configuration file fed into OpusDistillery can be found here.

How to run

from transformers import MarianMTModel, MarianTokenizer
model_name = "Helsinki-NLP/opus-mt_tiny_deu-eng"
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)
tok = tokenizer("Hallo, wie geht es dir?", return_tensors="pt").input_ids
output = model.generate(tok)[0]
tokenizer.decode(output, skip_special_tokens=True)

Benchmarks

Teacher

testset	BLEU	chr-F	COMET
Flores+	40.6	66.7	0.8559
Bouquet	42.5	63.5	0.8674

Student

testset	BLEU	chr-F	COMET
Flores+	37.6	64.4	0.8623
Bouquet	37.6	59.9	0.8556

Marian models

We also provide Marian-compatible versions of this model. To use them, compile Marian and run decoding with marian-decoder, for example:

marian-decoder \
  -i input.txt \
  -c final.model.npz.best-perplexity.npz.decoder.yml \
  -m final.model.npz.best-perplexity.npz \
  -v vocab.spm vocab.spm