ankekat1000/toxic-bert-german

Model description

This model is a fine-tuned version of the bert-base-german-cased model by deepset to classify toxic German-language user comments.

How to use

You can use the model with the following code.

#!pip install transformers

from transformers import AutoModelForSequenceClassification, AutoTokenizer, TextClassificationPipeline

model_path = "ankekat1000/toxic-bert-german"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForSequenceClassification.from_pretrained(model_path)

pipeline =  TextClassificationPipeline(model=model, tokenizer=tokenizer)
print(pipeline('du bist blöd.'))

You can apply the pipeline on a data set.

df['result'] = df['comment_text'].apply(lambda x: pipeline(x[:512])) #Cuts after max. legth of tokens for this model, which is 512 for this model.
# Afterwards, you can make two new columns out of the column "result", one including the label, one including the score.
df['toxic_label'] = df['result'].str[0].str['label']
df['score'] = df['result'].str[0].str['score']

Training

The pre-trained model bert-base-german-cased model by deepset was fine-tuned on a crowd-annotated data set of over 14,000 user comments that has been labeled for toxicity in a binary classification task.

As toxic, we defined comments that are inappropriate in whole or in part. By inappropriate, we mean comments that are rude, insulting, hateful, or otherwise make users feel disrespected.

Language model: bert-base-cased (~ 12GB)
Language: German
Labels: Toxicity (binary classification)
Training data: User comments posted to websites and facebook pages of German news media, user comments posted to online participation platforms (~ 14,000)
Labeling procedure: Crowd annotation
Batch size: 32
Epochs: 4
Max. tokens length: 512
Infrastructure: 1xGPU Quadro RTX 8000
Published: Oct 24th, 2023

Evaluation results

Accuracy:: 86%
Macro avg. f1:: 75%

Label	Precision	Recall	F1	Nr. comments in test set
not toxic	0.94	0.94	0.91	1094
toxic	0.68	0.53	0.59	274