训练到150k，但loss仍在0.5左右，attention图时好时坏是为什么？loss为什么减不下去？

下载的aidatatang_200zh.tgz数据集，在pre.py预处理之后，进行生成器训练。下面是部分图：


step=155000⬆
step=143000⬆

总觉得不对劲，这种情况是正常的吗？若不正常应该如何处理？

Start with the preprocessing flow in pre.py and the generator-training entry point, then review the reported loss and attention images around steps 143000 and 155000. Determine whether the plateau and changing attention are expected for the aidatatang_200zh.tgz training setup; done requires a reproducible diagnosis and a clearly identified corrective action if they are not.

Written by the indexing model from the issue text.