Trained a roberta-base bias classification model with Cleanlab preprocessing and class weighting. Training accuracy = 87.8% Test accuracy = 74.8%