About labels

[chikin-lau]

Is the target response x used for training in your paper the original target response or the extended target response? I see in the extended dataset you posted that the target response uses a retrieval style model. isn't compac a generative model?