german_toxicity_classifier_plus

This model was trained for toxicity labeling. Label_1 means TOXIC, Label_0 means NOT TOXIC

The model was fine-tuned based off the already existing sentiment classifier oliverguhr/german-sentiment-bert . The aforementioned classifier performed poorly (44% accuracy on my test sample), so I trained the current toxicity classifier. It was noted that the same performance achieved training on the dbmdz/bert-base-german-cased model.

The accuracy is 91% on the test split during training and 83% on a manually picked (and thus harder) sample of 200 sentences (100 label 1, 100 label 0) at the end of the training.

The model was finetuned on 37k sentences. The train data was the translations of the English data (around 30k sentences) from the multilingual_detox dataset by Skolkovo Institute using the opus-mt-en-de translation model by Helsinki-NLP and semi-manually collected data (around 7 k) by crawling the dict.cc web dictionary and the Reverso Context.