We present an efficient method for adapting a monolingual Large Language Model (LLM) to another language, addressing challenges of catastrophic forgetting and tokenizer limitations. We focus this study on adapting Llama 2 to Arabic. Our two-stage approach begins with expanding the vocabulary and training only the embeddings matrix, followed by full model continual pre-training on a bilingual corpus. By continually pre-training on a mix of Arabic and English corpora, the model retains its proficiency in English while acquiring capabilities in Arabic. Our approach results in significant improvements in Arabic and slight enhancements in English, demonstrating cost-effective cross-lingual transfer. We perform ablations on embedding initialization techniques, data mix ratios, and learning rates and release a detailed training recipe. To demonstrate generalizability of this approach we also adapted Llama 3 8B to Arabic and Llama 2 13B to Hindi.
@article{arxiv.2407.12869,
title = {Bilingual Adaptation of Monolingual Foundation Models},
author = {Gurpreet Gosal and Yishi Xu and Gokul Ramakrishnan and Rituraj Joshi and Avraham Sheinin and Zhiming and Chen and Biswajit Mishra and Natalia Vassilieva and Joel Hestness and Neha Sengupta and Sunil Kumar Sahu and Bokang Jia and Onkar Pandit and Satheesh Katipomu and Samta Kamboj and Samujjwal Ghosh and Rahul Pal and Parvez Mullah and Soundar Doraiswamy and Mohamed El Karim Chami and Preslav Nakov},
journal= {arXiv preprint arXiv:2407.12869},
year = {2024}
}