My solution of Kaggle competition: https://www.kaggle.com/competitions/lmsys-chatbot-arena
- I am only using 0.5% of the train data
- I will compute the embeddings using TPU in a different notebook (bert-embeddings-tpu.ipynb) and then load in the main notebook