There were some comments on Reddit suggesting that cutting the dataset at 15 and removing 40% of words was not the best move. I have locally built a version with the limit set to 30.

For the interested:
https://imgur.com/a/Rz7Cw6x