I think you can see this one of two ways: you could also consider it a miracle that the qwen models are able to perform so well when being trained on inefficient wrapper code data.