I feel excited to read papers focusing on multi-task learning for code-related tasks. Yet I noticed that the 5 training datasets and evaluation codes are not open-sourced, which makes the experimental results irreproducible. I believe this work would make an even bigger impact if it could be fully open-sourced.