Skip to content

fix: seq2seq decode returns primary choice (strip alternates) - #35

Merged
ronaldtse merged 1 commit into
mainfrom
fix/first-alternates-decode
Oct 11, 2026
Merged

ronaldtse merged 1 commit into
mainfrom
fix/first-alternates-decode

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Summary

  • Models trained on multi-reference corpora (QCRI silver) emit slash-separated alternate sets per token; the runtime returned them verbatim
  • Counting alternate-emitting tokens as single words scored them wrong: the AR news specialist corrects from 10.03 to 4.10 WER (8.95 to 1.56 DER) on WikiNews-2024 multiref under the same benchmark, first-alternate decode
  • first_alternates applied in Model.translate; pure-function tests; plane family unaffected

Test plan

  • pytest tests/test_first_alternates.py (3 tests, TDD red-green)
  • full suite: 86 passed, 1 xpassed

@ronaldtse
ronaldtse merged commit acdc35c into main Oct 11, 2026
3 of 6 checks passed
@ronaldtse
ronaldtse deleted the fix/first-alternates-decode branch October 11, 2026 08:29
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