Comment by gravypod

Since you know the size of the file beforehand you may be able to overfit some kind of text diffusion model instead of a transformer? May allow you to partially correct the model output using some other method and then fill in the blanks that were wrong from previous generations.