re trained the models on only tournament games and the ChalkSeedDiff

model with 10k epochs and 0.001 learning rate works much better than