quickmt-en-de Neural Machine Translation Model

quickmt-en-de is a reasonably fast and reasonably accurate neural machine translation model for translation from de into en.

Model Information

• Trained using eole
• 185M parameter transformer 'big' with 8 encoder layers and 2 decoder layers
• 20k separate Sentencepiece vocabularies
• Exported for fast inference to CTranslate2 format
• Training data: https://huggingface.co/datasets/quickmt/quickmt-train.de-en/tree/main

See the eole model configuration in this repository for further details and the eole-model for the raw eole (pytorch) model.

Usage with quickmt

You must install the Nvidia cuda toolkit first, if you want to do GPU inference.

Next, install the quickmt python library and download the model:

git clone https://github.com/quickmt/quickmt.git
pip install ./quickmt/

quickmt-model-download quickmt/quickmt-en-de ./quickmt-en-de

Finally use the model in python:

from quickmt import Translator

# Auto-detects GPU, set to "cpu" to force CPU inference
t = Translator("./quickmt-en-de/", device="auto")

# Translate - set beam size to 5 for higher quality (but slower speed)
sample_text = 'Dr. Ehud Ur, professor of medicine at Dalhousie University in Halifax, Nova Scotia and chair of the clinical and scientific division of the Canadian Diabetes Association cautioned that the research is still in its early days.'

t(sample_text, beam_size=5)

> 'Dr. Ehud Ur, Professor für Medizin an der Dalhousie University in Halifax, Nova Scotia und Vorsitzender der klinischen und wissenschaftlichen Abteilung der Canadian Diabetes Association warnte, dass die Forschung noch in ihren frühen Tagen ist.'

# Get alternative translations by sampling
# You can pass any cTranslate2 `translate_batch` arguments
t([sample_text], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)

The model is in ctranslate2 format, and the tokenizers are sentencepiece, so you can use ctranslate2 directly instead of through quickmt. It is also possible to get this model to work with e.g. LibreTranslate which also uses ctranslate2 and sentencepiece.

Metrics

bleu and chrf2 are calculated with sacrebleu on the Flores200 devtest test set ("deu_Latn"->"eng_Latn"). comet22 with the comet library and the default model. "Time (s)" is the time in seconds to translate (using ctranslate2) the flores-devtest dataset (1012 sentences) on an RTX 4070s GPU with batch size 32 (faster speed is possible using a large batch size).

	bleu	chrf2	comet22	Time (s)
quickmt/quickmt-en-de	40.17	66.26	86.91	1.1
Helsink-NLP/opus-mt-en-de	36.06	63.53	84.63	3.72
facebook/nllb-200-distilled-600M	35.72	63.41	86.65	26.93
facebook/nllb-200-distilled-1.3B	38.61	65.01	87.99	46.57
facebook/m2m100_418M	28.57	57.76	79.75	21.46
facebook/m2m100_1.2B	36.24	63.37	85.82	41.5

quickmt-en-de is the fastest and highest quality (except possibly for nllb-200-distilled-1.3B which has a higher comet22 score).