Module G: Attack Lab
Damage a piece of text on purpose, then rerun Modules A, B, C and F on the result. You see which methods hold up and which collapse, instead of taking anyone's word for it.
How it works and its limits
Synonym substitution, sentence reordering and truncation are simple structural perturbations. Paraphrase uses a real local T5 model.
In this project's measurements paraphrase was the least damaging of the four. That is a property of this specific small checkpoint, chosen for its size, not a refutation of the published result it is based on. Read the limitations.
Needed to test Module A as well
Results appear here: the attacked text, and every module’s reading before and after.