I have revised the dataset and training settings.

I think it captures the "40010" style better than V1.0, though it has become slightly less stable as a result; please use negative prompts or adjust the weight to control the output