Cross-Lingual Supervision improves Large Language Models Pre-training (2305.11778v1)

Abstract: The recent rapid progress in pre-training LLMs has relied on using self-supervised LLMing objectives like next token prediction or span corruption. On the other hand, Machine Translation Systems are mostly trained using cross-lingual supervision that requires aligned data between source and target languages. We demonstrate that pre-training LLMs on a mixture of a self-supervised LLMing objective and the supervised Machine Translation objective, therefore including cross-lingual parallel data during pre-training, yields models with better in-context learning abilities. As pre-training is a very resource-intensive process and a grid search on the best mixing ratio between the two objectives is prohibitively expensive, we propose a simple yet effective strategy to learn it during pre-training.