# 2nd place solution

This is the github link for my notebook. The code I used is structured in various modules and scripts, but I had to copy paste everything into one notebook for submission, so I don't actually know if the notebook runs, but it should contain everything necessary to replicate my result.
[https://github.com/ndrwhoang/kaggle_comp_congress_tweet](https://github.com/ndrwhoang/kaggle_comp_congress_tweet)
I did the transformer + prediction head thing and went with the kitchen sink approach of feature engineering. Based on the similar results on the LB, I suspect others also somewhat follow this approach; what I have done more is gradual unfreezing and data leakage abuse (there are some overlap between test set and train set, you can get around +0.009 accuracy, though that depends on your original accuracy)