Steering Large Language Models for Machine Translation with Finetuning and In-Context Learning

This repository provides the code for the paper "Steering Large Language Models for Machine Translation with Finetuning and In-Context Learning". It was based on the lit-llama repository.

Quick Installation

To install run:

pip install -r requirements.txt
pip install -e .

Convert LLaMA checkpoint

Before fine-tuning the pre-trained model, you need to convert the original LLaMA checkpoint to be compatible with the tool. To do this, run:

python translation_llm/convert_checkpoint.py \
    --output_dir <directory to save converted checkpoint> \
    --checkpoint_dir <directory with original checkpoint> \
    --model_size <7B/13B>

Finetune

Look at the finetune_lora.sh and finetune_no_lora.sh scripts for examples of how to finetune the models.

Generate

Look at the generate_lora.sh and generate_no_lora.sh scripts for examples of how to generate from the models.

Evaluate

For the evaluation environment, you need to install the packages sacrebleu and unbabel-comet.

Look at the eval.sh script for an example on how to evaluate the models.