Hi all, this is a very interesting and timely paper targeting at an important problem. I'd like to discuss a bit about the paper objective and the training setup.

Paper Objective

Assuming that we have an "original program" $p_o$ and its transformed version $p_t$, the goal of this paper is to ask the code owner to release $p_t$ (maybe to Github) so that even $p_t$ is used for training a model, this model cannot complete the content in $p_o$.

Could you please let me know whether I understand in the correct way?

Training setup

As the paper evaluates "self-completion," can I assume that you don't have to split POJ-104 into train/dev/test datasets?

Let us denote the original training data set $D_o$ (i.e., the whole POJ-104 dataset), and the transformed datasets as $D_{UE}$ and $D_{B2}$.

If I understand correctly, to obtain results in Table V, you use $D_o$ to evaluate the model trained on $D_{UE}$ and $D_{B2}$. Right?