zoomin·2 lata temu·discussAn eval on 25% of the eval dataset is fishy? Why not 100%? Are they training with the eval set? Also that dataset is almost all python.