I have used your pretrained model to generate summaries for giga word , duc2003 and duc2004. I would like to know what is the pretrained model trained on and how it is generating only titles for the giga word corpus and summaries for the other two datasets?