519. Feedback Prize - English Language Learning | feedback-prize-english-language-learning
首先,感谢主办方举办这次比赛。
我在这次比赛中学到了很多关于NLP的知识。
祝贺获奖者们!
| 模型 | CV得分 |
|---|---|
| (*1,4,5) deberta-v3-base (伪标签 fb1,2) | 0.4528044 |
| (*4,5) deberta-v3-large | 0.4568590 |
| (*2,4,5) deberta-v3-base (主题) | 0.4537212 |
| (*3,4,5) deberta-v3-base (原始损失) | 0.4564600 |
※ (*1,4,5) 使用了「伪标签」、「拼接最后四层隐藏层」和「Dropout」。
self.mse = nn.MSELoss(reduction='none')
l = self.mse(y_pred, y_true) + eps # eps=1e-9
loss = torch.mul(torch.pow(l,alpha),torch.sqrt(l))
# soft→alpha=1, hard→alpha=2
self.dropouts = nn.ModuleList([nn.Dropout(0.2) for _ in range(5)])
output = sum([self.fc(dropout(feature)) for dropout in self.dropouts])/5