21. ● Diverse pre-trained embeddings (baseline public LB of 0.9877)
● Translations as train/test-time augmentation (TTA) (boosted LB from 0.9877 to
0.9880)
● Rough-bore pseudo-labelling (PL) (boosted LB from 0.9880 to 0.9885)
● Robust CV + stacking framework (boosted LB from 0.9885 to 0.9890)
上位陣の解法: 1st place solution
21