🌍 CrossIn

An Efficient Instruction Tuning Approach for Cross-Lingual Knowledge Alignment

📋 Overview

CrossIn is a novel approach for efficient instruction tuning that focuses on cross-lingual knowledge alignment. This repository contains the official implementation of our paper "CrossIn: An Efficient Instruction Tuning Approach for Cross-Lingual Knowledge Alignment".

👥 Authors

• Geyu Lin
• Bin Wang
• Zhengyuan Liu
• Nancy F. Chen

🚀 Getting Started

Prerequisites

Install the required dependencies:

pip install -r requirements.txt

Usage Guide

1️⃣ Build CrossIn Data

bash sample_scripts/build_data.sh

Note: You need to download the Alpaca and Platypus datasets into the data/ folder first.

2️⃣ Training

bash sample_scripts/run_training.sh <dataset_name> <stage> <exp_group> <prompt> <batch> <epoch> <lr>

3️⃣ Evaluation

Evaluation is performed using the SeaEval framework.

📚 Citation

If you find this work useful, please consider citing our paper:

@misc{lin2024crossinefficientinstructiontuning,
    title={CrossIn: An Efficient Instruction Tuning Approach for Cross-Lingual Knowledge Alignment}, 
    author={Geyu Lin and Bin Wang and Zhengyuan Liu and Nancy F. Chen},
    year={2024},
    eprint={2404.11932},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2404.11932}, 
}