> The mentioned approach is fundamentally flawed, since the inputs are used during pretraining constituting to a leakage, a universally recognized flaw of ML training.

I saw this on the community note for the last blog you wrote - anything to do here.