Kaggle Unintended Bias in Toxicity Classification Challenge: https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification
Current ensemble model ranks 377th on Private LB with custom Bias AUC score of 0.93832 against Private LB highest of 0.94734
Ensemble Details:
- BiLSTM model with 5-fold CV trained using FastText + GloVe embeddings (300d each)
- Fine-tuned BERT baseline
Potential Areas of Improvement:
- Finetune BERT baseline over entire training set (present limit is 50% due to computational constraints)
- Try ensembling different BERT baselines (cased / uncased)