Will the effect of this pluggable component be affected by data contamination?

[TinderZ]

As the title suggests, some experiments in the paper were conducted on Qwen2.5, but existing research indicates that models in this family may suffer from data contamination.

Could you conduct experiments on more model families, especially those known to be "clean," as this might lead to even greater improvements in metrics?

[TinderZ]

I think this is a relatively reasonable concern. If the model has seen these types of questions during training, its ability to "help the model better align with and follow the given instruction" might decrease.