kazRush-kk-ru
kazRush-kk-ru is a translation model for translating from Kazakh to Russian. The model was trained with randomly initialized weights based on the T5 configuration on the available open-source parallel data.
Usage
Using the model requires sentencepiece library to be installed.
After installing necessary dependencies the model can be run with the following code:
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import torch
device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModelForSeq2SeqLM.from_pretrained('deepvk/kazRush-kk-ru').to(device)
tokenizer = AutoTokenizer.from_pretrained('deepvk/kazRush-kk-ru')
@torch.inference_mode
def generate(text, **kwargs):
inputs = tokenizer(text, return_tensors='pt').to(device)
hypotheses = model.generate(**inputs, num_beams=5, **kwargs)
return tokenizer.decode(hypotheses[0], skip_special_tokens=True)
print(generate("Анам жақтауды жуды."))
You can also access the model via pipeline wrapper:
>>> from transformers import pipeline
>>> pipe = pipeline(model="deepvk/kazRush-kk-ru")
>>> pipe("Иттерді кім шығарды?")
[{'translation_text': 'Кто выпустил собак?'}]
Data and Training
This model was trained on the following data (Russian-Kazakh language pairs):
| Dataset | Number of pairs |
|---|---|
| OPUS Corpora | 718K |
| kazparc | 2,150K |
| wmt19 dataset | 5,063K |
| TIL dataset | 4,403K |
Preprocessing of the data included:
- deduplication
- removing trash symbols, special tags, multiple whitespaces etc. from texts
- removing texts that were not in Russian or Kazakh (language detection was made via facebook/fasttext-language-identification)
- removing pairs that had low alingment score (comparison was performed via sentence-transformers/LaBSE)
- filtering the data using opusfilter tools
The model was trained for 56 hours on 2 GPUs NVIDIA A100 80 Gb.
Evaluation
Current model was compared to another open-source translation model, NLLB. We compared our model to all version of NLLB, excluding nllb-moe-54b due to its size.
The metrics - BLEU, chrF and COMET - were calculated on devtest part of FLORES+ evaluation benchmark, most recent evaluation benchmark for multilingual machine translation.
Calculation of BLEU and chrF follows the standart implementation from sacreBLEU, and COMET is calculated using default model described in COMET repository.
| Model | Size | BLEU | chrf | COMET |
|---|---|---|---|---|
| nllb-200-distilled-600M | 600M | 18.0 | 47.3 | 85.6 |
| This model | 197M | 18.8 | 48.7 | 86.7 |
| nllb-200-1.3B | 1.3B | 20.4 | 49.3 | 87.9 |
| nllb-200-distilled-1.3B | 1.3B | 20.8 | 49.6 | 88.1 |
| nllb-200-3.3B | 3.3B | 21.5 | 50.7 | 88.7 |
Examples of usage:
>>> print(generate("Балық көбінесе сулардағы токсиндердің жоғары концентрацияларына байланысты өледі."))
Рыба часто умирает из-за высоких концентраций токсинов в воде.
>>> print(generate("Өткен 3 айда 80-нен астам қамалушы ресми түрде айып тағылмастан изолятордан шығарылды."))
За прошедшие 3 месяца более 80 арестованных были официально извлечены из изолятора без обвинения.
>>> print(generate("Бұл тастардың он бесі өткен шілде айындағы метеориттік жаңбырға жатқызылады."))
Пятнадцать этих камней относят к метеоритным дождям прошлого июля.
Citations
@misc{deepvk2024kazRushkkru,
title={kazRush-kk-ru: translation model from Kazakh to Russian},
author={Lebedeva, Anna and Sokolov, Andrey},
url={https://huggingface.co/deepvk/kazRush-kk-ru},
publisher={Hugging Face},
year={2024},
}