Skip to content

Document attacking translation models with morpheus/seq2sick - #840

Merged
qiyanjun merged 1 commit into
masterfrom
docs/content-accuracy-fixes
Aug 15, 2026
Merged

Document attacking translation models with morpheus/seq2sick#840
qiyanjun merged 1 commit into
masterfrom
docs/content-accuracy-fixes

Conversation

@qiyanjun

Copy link
Copy Markdown
Member

Summary

  • Add a verified, working MorpheusTan2020 example for attacking a seq2seq translation model (BLEU vs. a reference translation)
  • Cross-reference the equivalent Seq2SickCheng2018BlackBox variant (compares against the model's own unperturbed output instead of a reference)
  • Answers the open question in How to use textattack on generative models?聽#725 about attacking BART on an en-de translation task, now that HuggingFaceModelWrapper's .generate() path (Issues attacking se2seq model (seq2sick)聽#771) makes it possible

Test plan

  • Ran the MorpheusTan2020 example end-to-end against t5-small (translation + attack succeeded)
  • Verified the .generate() code path against a BART-family checkpoint separately
  • make docs rebuild introduces no new Sphinx warnings (same 4 pre-existing, unrelated ones)

Closes/relates to #725

Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com

Add a verified, working example for MorpheusTan2020 attacking a
seq2seq translation model (BLEU vs. a reference translation), and
cross-reference the equivalent seq2sick variant (compares against the
model's own unperturbed output instead). Answers the open question in
#725 about attacking BART on
an en-de translation task, now that HuggingFaceModelWrapper's .generate()
path (#771) makes it possible.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
@qiyanjun
qiyanjun merged commit f800c2c into master Aug 15, 2026
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant