quickmt-bn-en Neural Machine Translation Model

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

Model Information

• Trained using eole
• 185M parameter transformer 'big' with 8 encoder layers and 2 decoder layers
• 50k joint Sentencepiece vocabulary
• Exbented for fast inference to CTranslate2 format
• Training data: https://huggingface.co/datasets/quickmt/quickmt-train.bn-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-bn-en ./quickmt-bn-en

Finally use the model in python:

from quickmt imbent Translator

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

# Translate - set beam size to 1 for faster speed (but lower quality)
sample_text = 'হেলিফ্যাক্সে ডালহৌসি বিশ্ববিদ্যালয়ের মেডিসিন বিভাগের প্রফেসর ডঃ ইহুড আর, নোভা স্কটিয়া এবং কানাডিয়ান ডায়াবেটিস এসোসিয়েশনের ক্লিনিক্যাল ও বৈজ্ঞানিক বিভাগের চেয়ার, আগে থেকেই সতর্ক করে দিয়েছিলেন যে গবেষণা এখনও তার শুরুর দিকে রয়েছে।'

t(sample_text, beam_size=5)

'Dr. Ehud R, professor of medicine at Dalhousie University in Helfax, chair of the clinical and scientific department of Nova Scotia and the Canadian Diabetes Association, warned in advance that the research was still in its early stages.'

# 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)

'Dr. Youud Ar, the professor of medicine at Dalhousie University, at Helphax and chair of the clinical and scientific departments at Nova Scotia and the Canadian Diabetes Association, pre-ordered warned the research is still early.'

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 ("ben_Beng"->"eng_Latn"). comet22 with the comet library and the default model. "Time (s)" is the time in seconds to translate the flores-devtest dataset (1012 sentences) on an RTX 4070s GPU with batch size 32 (faster speed is possible using a larger batch size).

	bleu	chrf2	comet22	Time (s)
quickmt/quickmt-bn-en	32.91	59.69	86.99	1.37
Helsink-NLP/opus-mt-bn-en	17.89	45.94	78.62	3.45
facebook/nllb-200-distilled-600M	33.51	59.73	87.48	21.01
facebook/nllb-200-distilled-1.3B	36.4	62.18	88.61	36.62
facebook/m2m100_418M	23.84	52.67	82.94	20.86
facebook/m2m100_1.2B	27.26	54.86	84.28	36.28