I wonder if this is a case where you want an encoder-decoder model. It seems very much like a translation task, only one where training data is embarrassingly easy to synthesize by just grabbing sentences from a corpus and occasionally swapping, inserting, and deleting characters.

In terms of attention masking, it seems like you want the input to be unmasked, since the input is fixed for a given “translation”, and then for the output tokens to use causally masked self attention plus cross attention with the input.

I wonder if you could get away with a much smaller network this way because you’re not pointlessly masking input attention for a performance benefit that doesn’t matter.