YaRN is a compute-efficient method to extend the context window of transformer-based language models, allowing them to effectively utilize and extrapolate to longer context lengths. It surpasses previous methods and can extrapolate beyond the limited context of a fine-tuning dataset.