Tag
This paper presents a synthetic training pipeline that uses perturbed public documents to generate context-dependent training data, significantly improving large language models' performance on context learning tasks like CL-bench without human annotation.