LLM_BTI

Enhancing Multilingual Translation Utilizing Back-Translation During Inference

Overview

This project implements advanced multilingual translation using a combination of LLM-based generation, back-translation, and iterative refinement. It utilizes models such as NLLB (facebook/nllb-200-distilled-600M) and Llama3 (via Portkey), and evaluates results using metrics like COMET and chrF.

Features

• Back-Translation during Inference for robust evaluation and improvement of translations
• Convergent iterative refinement with LLM-based critique and correction
• Multiple translation models (NLLB, Llama3)
• Evaluation using standard metrics: COMET, chrF
• Example pipelines and recipe for Flores200 dataset evaluation

Requirements

See requirements.txt; install with:

pip install -r requirements.txt

Required Model/Service Keys:
Create a .env file with your keys for:

• PORTKEY_API_KEY
• GROQ_VIRTUAL_KEY
• HUGGING_FACE_HUB_TOKEN

Usage

Back-Translation Pipeline

Run the main script for translation, critique, and evaluation:

python LLM_BTI.py

For a free prompt demo with user input:

python LLM_BTI_free_prompt.py

Evaluation Pipeline

The pipeline can be used to automatically evaluate the translation quality on the Flores200 dataset (or other supported datasets). See classes and methods in LLM_BTI.py for batch processing and result plotting.

File Structure

• LLM_BTI.py — main script for dataset-based translation and evaluation
• LLM_BTI_free_prompt.py — console-based user prompt workflow

Citation

If you use this code/project in your research, please cite

@software{jamesjyoon_LLM_BTI,
  author = {James Y. Yoon},
  title = {LLM_BTI: Enhancing Multilingual Translation Utilizing Back-Translation During Inference},
  year = {2025},
  url = {https://github.com/jamesjyoon/LLM_BTI}
}

***

Let me know if you want any customization to the requirements or the README structure!

[1](https://github.com/jamesjyoon/LLM_BTI)