It was the 30s... The creators of Talkie-1930 explicitly admit that their filtering of the pre-training corpus was imperfect. Nobody has dug validated the training data fully, it may have had leaks. It would be intresting if this happened. I hope one day we can peer-review validate this experiment
