Skip to content

Generate a 1 million-sample warm-up dataset for neural machine translation from a 700 million-word Mongolian text corpus using the Google Translate service

Notifications You must be signed in to change notification settings

sharavsambuu/english-mongolian-nmt-dataset-augmentation

Repository files navigation

Tasks

Main goal is to improve existing translation pairs using Llama 3.1 and some prompts

  1. Find a way to call LM Studio using Python
  2. Prepare a prompt to detect badly translated pairs
  3. Prepare a prompt to improve mongolian to english translation pair
  4. Publish improved translations
  5. Train on improved translation dataset on the Colab and Tensorflow 2`

Notes

700 million word Mongolian news data set https://github.com/tugstugi/mongolian-bert

Python library for google translation, textblob https://textblob.readthedocs.io/en/dev/

Steps

sudo apt install parallel tor build-essential rar virtualenv bsdmainutils

virtualenv -p python3 env && source env/bin/activate && pip install -r requirements.txt

./install_translate.sh

python3 datasets/dl_and_preprop_mn_news.py

./generate_sentences.sh && ./criterion_sentences.sh && ./prepare_translation.sh && ./split_sentences.sh

./runtask.sh

References and links

https://lmstudio.ai/

https://huggingface.co/lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF

https://medium.com/@LakshmiNarayana_U/lm-studios-python-play-crafting-code-creatively-86c21d18d34a

https://opensource.com/article/18/5/gnu-parallel

https://askubuntu.com/questions/147241/execute-sudo-without-password

Released datasets

[2019/09/10] 5K unvalidated sentences of pairs for english to mongolian.

https://gist.github.com/sharavsambuu/be9001ddcb954565606466a3556bbf27

[2019/09/13] 94K unvalidated sentences of pairs for english to mongolian.

https://drive.google.com/file/d/1GNo1XJxRFxjey5VDsHjLvj9upXJOqd3e/view?usp=sharing

[2019/10/10] 1 million mongolian to english sentence pairs.

https://drive.google.com/file/d/14AtTVgibirSdHYTBFM9G1XPS7DvM5SdE/view?usp=sharing

Neural machine translation colab experiment, based on official tensorflow transformer tutorial

Pretrained version for fast inference Inference pretrained version

Train on Colab Train In Colab

About

Generate a 1 million-sample warm-up dataset for neural machine translation from a 700 million-word Mongolian text corpus using the Google Translate service

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published