Demonstration that finetuning RoPE model on larger sequences than the pre-trained model adapts the model context limit

[EMNLP 2023] Adapting Language Models to Compress Long Contexts