Why I use the same code and dataset to train the model,then use the context from train set to generate QA pairs, but they are not as good as the train set(eg. only one or two is good in four generated QA pairs)?

[CoderJoyce]

[seanie12]

hi, how did you generate the QA pairs?
If you sample multiple latent variables, the quality of some generated QA pairs are not good. So we refine them with QA model

[CoderJoyce]

I ran the code generate_qa.py to generate QA pairs.
How can I refine them with QA model? Is there any code available?