Repository navigation
How to get the task name of the victim model during the transformations #444
Description
Hi. First of all, thank you for your awesome project!
I'm trying to add a new recipe with MLM-based substitution (e.g., BAE or CLARE).
And I've struggled to load the BERT that is also trained on the same dataset with the victim model.
For instance, if the victim model is some arbitrary model fine-tuned on the YELP polarity dataset, I want to use the BERT that is also trained on YELP for word substitution or insertion.
To this end, It seems that my recipe class should be able to know the information about the victim model (maybe somewhere like the below code, where the recipe class calls the transformation class).
https://github.com/QData/TextAttack/blob/master/textattack/attack_recipes/bae_garg_2019.py#L53
but I'm not sure whether there exists any shortcut for this problem. Could you give me any advice for my purpose?
Thank you!
Activity
donggrant commentedon Oct 29, 2021
This is also an enhancement that I think could be very useful. When trying to enhance Tutorial 9, I needed to find a faster alternative to the CamemBert model that was pretrained on the Allocine dataset. It was hard to locate an alternative dataset where the CamemBert model was pretrained on.
qiyanjun commentedon Aug 13, 2026
This is a real API design gap — there's currently no clean way for a
Transformationto know what dataset/task the victim model was trained on, so MLM-based transformations (BAE/CLARE) can't pick a same-domain masked-LM the way this issue and @donggrant's comment both want. Threading that metadata fromAttackRecipe.build(model_wrapper)down intoTransformation.__init__would need a design decision on where that metadata lives (onModelWrapper? passed explicitly tobuild()?) — not something to guess at unilaterally. Leaving open as a real, still-valid feature request.