Instruct Multilingual

This repository contains code to translate datasets into multiple languages using the NLLB (No Language Left Behind) model. The model can be used:

• From a server to translate a single text or entire datasets.
• Using the translate script to translate a single text.

Note: This repository has been tested on a Linux machine using a Nvidia GPU. The code assumes access to a GPU. Depending on your hardware, you might need to modify the code to change the number of GPUs and the batch size.

Setup

It is recommended to use a virtual environment to install the dependencies.

conda create -n instructmultilingual python=3.8.10 -y
conda activate instructmultilingual
pip install -r requirements.txt

Inference Server

The inference server is a FastAPI application that can be used to translate a single text or entire datasets.

Convert the models first

For efficient inference, the model is converted using CTranslate2.

mkdir models
ct2-transformers-converter --model facebook/nllb-200-3.3B --output_dir models/nllb-200-3.3B-converted

Run the server locally

To start the server, we need to run the following command:

uvicorn instructmultilingual.server:app --host 0.0.0.0 --port 8000

Using docker to run the server

To run the server using docker, we need to build and run the docker image, and run the server.

# Build
docker build -t instruct-multilingual .

# Run
docker run -it --rm --gpus 1 -p 8000:8000 -v $(pwd):/instruct-multilingual instruct-multilingual

Client Side

This script translate the samsum dataset using the inference server. It showcases how to use the inference server to translate a single text and entire datasets.

python main.py

Translate

We also provide a script to translate a single text from the CLI. This script downloads the model from Hugging Face and translates the text provided into the target language.

python -m instructmultilingual.translate \
          --text="Cohere For AI will make the best instruct multilingual model in the world" \
          --source_language="English" \
          --target_language="Egyptian Arabic"

Translate an instructional dataset from xP3 (or any dataset repo from HuggingFace Hub)

An example of using translate_dataset_from_huggingface_hub to translate PIQA with the finish_sentence_with_correct_choice template into languages used by Multilingual T5 (mT5) model

from instructmultilingual.translate_datasets import translate_dataset_from_huggingface_hub

translate_dataset_from_huggingface_hub(
    repo_id = "bigscience/xP3",
    train_set = ["en/xp3_piqa_None_train_finish_sentence_with_correct_choice.jsonl"],
    validation_set = ["en/xp3_piqa_None_validation_finish_sentence_with_correct_choice.jsonl"],
    test_set = [],
    dataset_name="PIQA",
    template_name="finish_sentence_with_correct_choice",
    splits=["train", "validation"],
    translate_keys=["inputs", "targets"],
    url= "http://localhost:8000/translate",
    output_dir= "/home/weiyi/instruct-multilingual/datasets",
    source_language= "English",
    checkpoint="facebook/nllb-200-3.3B",
)