Built a ML model using an ensemble of BERT (large and small) and LSTM-based language models with different loss functions to identify toxicity in online conversations, where toxicity is defined as anything rude, disrespectful or otherwise likely to make someone leave a discussion.
Computed 4 folds for each of the five models and final inference had to be done in Kaggle kernel in under 2 hours according to the competiton rules.