we examine and ensure that the knowledge in the responses is not contained within the model, and eliminate responses that might lead to memorization;

[real-brilliant]

YI在构建SFT数据时，针对幻觉问题进行如下优化：

to reduce hallucinations, we examine and ensure that the knowledge in the responses is not contained within the model, and
eliminate responses that might lead to memorization;

这如何理解呢？确保LLM基于且仅基于instruction所给出的knowledge进行生成吗