Repository navigation
BackTranslationAugmenter produces an incorrect number of transformations #800
Description
Describe the bug
Calling the augment function of a BackTranslationAugmenter produces a list with a length that appears to be non-deterministic, sometimes not matching the value of the transformations_per_example initialization parameter.
To Reproduce
Run the following code:
import textattack
augmenter = textattack.augmentation.recipes.BackTranslationAugmenter(transformations_per_example=5)
s = "You are in the kitchen."
ls = augmenter.augment(s)
print(len(ls))
Expected behavior
Expected output: 5
Actual output: 3
System Information (please complete the following information):
- OS: MacOS Sonoma 14.6.1
- Library versions: torch==2.3.1, transformers==4.42.4
- Textattack version: textattack==0.3.10
- Full
pip freezeoutput: https://pastebin.com/VPt7JdRY
Activity
qiyanjun commentedon Aug 13, 2026
Root cause:
Augmenter.augmentaccumulates results in aset()but only ever attempts exactlytransformations_per_exampleouter-loop calls to the transformation.BackTranslationAugmenterdefaults tochained_back_translation=5(randomly chained through 5 languages), and for a short sentence like yours, two of those attempts can easily land on the identical final phrasing — the set silently collapses the duplicate, so you get fewer results than requested. It's not really about back-translation specifically; any transformation with limited output diversity for a given input hits this.Opened #831 with a bounded retry (up to 3x the requested count) so the loop backfills collisions instead of accepting a short result. Transformations with large search spaces (the common case) see zero behavior change. Will close automatically once that merges.
Fix remaining review findings from PR QData#831
Fix six more bugs from the unresolved-bugs triage list