A question about Wizardcoder paper

[Smith-xuan]

Dear author, hello! I have a question about the WizardCoder paper.
According to my understanding from the paper, after each round of evolution, you merge the data from that round with the data from previous rounds for fine-tuning, resulting in a performance iteration growth effect similar to Figure 3 in the paper.

However, since the evolution of the dataset is independent of the fine-tuning process, why not merge the data from each round together and perform fine-tuning only once? For example, if the original data used for fine-tuning is (0), (0,1), (0,1,2), (0,1,2,3), why not directly use only the data from (0,1,2,3) for fine-tuning? Does this iterative fine-tuning approach lead to performance improvements in the model? I would like to ask if your team has conducted experiments in this regard.
If you are willing to provide an answer, I would be extremely grateful！