Most Transformer language models are primarily pretrained on English text, limiting their use for other languages. As the model sizes grow, the performance gap between English and other languages with fewer compute and data resources increases even further. Consequently, more resource-efficient training methods are needed to bridge the gap for lang...

Efficient Language Model Training through Cross-Lingual and Progressive Transfer Learning