Convulational neural networks can produce texts/codes, can this? Id est, train with inputs of pairs of [github source codes, descriptions], to give new descriptions to output new source codes?

[ETERNALBLUEbullrun]

The input is vectors of floats/integers, so just convert sources/texts to integers/hex, plus convert descriptions to integers/hex, plus convert outputs from integers/hex to texts