---
license: mit
model-index:
- name: xlm-roberta-base-offensive-text-detection-da
  results: []
widget:
- text: "Din store idiot"
---

# Danish Offensive Text Detection based on ELECTRA-small

This model is a fine-tuned version of [xlm-roberta-base](https://huggingface.co/xlm-roberta-base) on a dataset consisting of approximately 5 million Facebook comments on [DR](https://dr.dk/)'s public Facebook pages. The labels have been automatically generated using weak supervision, based on the [Snorkel](https://www.snorkel.org/) framework.

The model achieves second place on a test set consisting of 500 Facebook comments annotated by two people, of which 41.2% were labelled as offensive:

| **Model** | **Precision** | **Recall** | **F1-score** |
| :-------- | :------------ | :--------- | :----------- |
| [`alexandrainst/electra-small-offensive-text-detection-da`](https://huggingface.co/alexandrainst/electra-small-offensive-text-detection-da) | 85.45% | 91.26% | **88.26%** |
| `alexandrainst/xlm-roberta-base-offensive-text-detection-da` (this) | 83.48% | **93.20%** | 88.07% |
| [`A-ttack`](https://github.com/ogtal/A-ttack) | **99.17%** | 58.25% | 73.39% |
| [`DaNLP/da-electra-hatespeech-detection`](https://huggingface.co/DaNLP/da-electra-hatespeech-detection) | 92.19% | 57.28% | 70.66% |
| [`Guscode/DKbert-hatespeech-detection`](https://huggingface.co/Guscode/DKbert-hatespeech-detection) | 84.91% | 43.69% | 57.69% |

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 32
- eval_batch_size: 32
- gradient_accumulation_steps: 1
- total_train_batch_size: 32
- seed: 4242
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- max_steps: 500000
- fp16: True
- eval_steps: 1000
- early_stopping_patience: 100


### Framework versions

- Transformers 4.20.1
- Pytorch 1.11.0+cu113
- Datasets 2.3.2
- Tokenizers 0.12.1