PoSE: Efficient Context Window Extension of LLMs via Positional Skip-wise Training

Publication
arXiv preprint arXiv:2309.10400