---
language:
- de
tags:
- distilbert
- german
- classification
datasets:
- germeval21
widget:
- text: "Das ist ein guter Punkt, so hatte ich das noch nicht betrachtet."
  example_title: "Agreement (non-toxic)"
- text: "Wow, was ein geiles Spiel. Glückwunsch."
  example_title: "Football (non-toxic)"
- text: "Halt deine scheiß Fresse, du Arschloch"
  example_title: "Silence (toxic)"
- text: "Verpiss dich, du dreckiger Hurensohn."
  example_title: "Dismiss (toxic)"
---

# German Toxic Comment Classification

## Model Description

This model was created with the purpose to detect toxic or potentially harmful comments.

For this model, we fine-tuned a German DistilBERT model [distilbert-base-german-cased](https://huggingface.co/distilbert-base-german-cased) on a combination of five German datasets containing toxicity, profanity, offensive, or hate speech.


## Intended Uses & Limitations

This model can be used to detect toxicity in German comments.
However, the definition of toxicity is vague and the model might not be able to detect all instances of toxicity.

It will not be able to detect toxicity in languages other than German.


## How to Use

```python
from transformers import pipeline

model_hub_url = 'https://huggingface.co/ml6team/distilbert-base-german-cased-toxic-comments'
model_name = 'ml6team/distilbert-base-german-cased-toxic-comments'

toxicity_pipeline = pipeline('text-classification', model=model_name, tokenizer=model_name)

comment = "Ein harmloses Beispiel"
result = toxicity_pipeline(comment)[0]
print(f"Comment: {comment}\nLabel: {result['label']}, score: {result['score']}")
```


## Limitations and Bias

The model was trained on a combinations of datasets that contain examples gathered from different social networks and internet communities. This only represents a narrow subset of possible instances of toxicity and instances in other domains might not be detected reliably.


## Training Data

The training dataset combines the following five datasets:

* GermEval18 [[dataset](https://github.com/uds-lsv/GermEval-2018-Data)]
    * Labels: abuse, profanity, toxicity
* GermEval21 [[dataset](https://github.com/germeval2021toxic/SharedTask/tree/main/Data%20Sets)]
    * Labels: toxicity
* IWG Hatespeech dataset [[paper](https://arxiv.org/pdf/1701.08118.pdf), [dataset](https://github.com/UCSM-DUE/IWG_hatespeech_public)]
    * Labels: hate speech
* Detecting Offensive Statements Towards Foreigners in Social Media (2017) by Breitschneider and Peters [[dataset](http://ub-web.de/research/)]
    * Labels: hate
* HASOC: 2019 Hate Speech and Offensive Content [[dataset](https://hasocfire.github.io/hasoc/2019/index.html)]
    * Labels: offensive, profanity, hate

The datasets contains different labels ranging from profanity, over hate speech to toxicity. In the combined dataset these labels were subsumed as `toxic` and `non-toxic` and contains 23,515 examples in total.

Note that the datasets vary substantially in the number of examples.


## Training Procedure

The training and test set were created using either the predefined train/test splits where available and otherwise 80% of the examples for training and 20% for testing. This resulted in in 17,072 training examples and 6,443 test examples.

The model was trained for 2 epochs with the following arguments:

```python
training_args = TrainingArguments(
    per_device_train_batch_size=batch_size,
    per_device_eval_batch_size=batch_size,
    num_train_epochs=2,
    evaluation_strategy="steps",
    logging_strategy="steps",
    logging_steps=100,
    save_total_limit=5,
    learning_rate=2e-5,
    weight_decay=0.01,
    metric_for_best_model='accuracy',
    load_best_model_at_end=True
)
```

## Evaluation Results

Model evaluation was done on 1/10th of the dataset, which served as the test dataset.

| Accuracy | F1 Score |  Recall  |  Precision  |
| -------- | -------- | -------- | ----------- |
|    78.50 |    50.34 |    39.22 |       70.27 |