That's kind of an interesting issue, I wonder if different tokenization would help. Like maybe putting a space between GTK and the number would put them in separate tokens and give better output.

More generally, do text AI's not support weighting terms like the image AI's do? Over in Stable Diffusion that sounds like something where I'd add a weight like "How do I create a <GTK3:1.2> interface in <PHP:1.1>?"