Tag
EndPrompt proposes a method for extending the context window of large language models using only short training sequences, by anchoring a terminal prompt with target-length positional indices. It achieves strong benchmark results with substantially less computation than full-length fine-tuning.