how is the adaption performance evaluated?

[wuyuhui-zju]

Hello, this is a great project! There was one detail I didn't fully understand when I read the paper. In section 2.5.2, " Among these downstream tasks, we thenconducted a number of adaptation experiments. For instance, we trained the refined model for task A on task B until the mod-els converged (e.g., A→B, stability→fluorescence). To determinewhether the adaptive performance was improving or degrading,it was compared to the initial performance for task A. The relationship between the intertask distribution distances and the mutual adaptability of the tasks was then calculated", the model was trained with A, then fine-tuned with B, and its performance on B's test set was compared to what? "it was compared to the initial performance for task A", did you use the model trained on A to predict B's test set? I think the comparison should be direct training with B, right? Thank you!