Questions Regarding the Nougat Model's Pre-training Process

[JasonKitty]

Dear Authors,

I have a couple of questions regarding the Nougat model as described in your paper.

Firstly, the paper mentions that the Nougat model employs a similar architecture to Donut, utilizing a Swin Transformer encoder and a Bart decoder. However, Donut uses a more complex pre-training technique, where the model is first taught to recognize text and then to handle structure. Could you clarify if the Nougat model also undergoes a similar pre-training process? It doesn't seem to be mentioned in the paper.

Secondly, could you please specify whether the Nougat model training started from the original pre-trained weights of swin_base_patch4_window12_384 and mbart-large-50, or did it begin with the pre-trained weights from the Donut model?

Thank you.