esceptico / toxic

End-to-end toxic Russian comment classification

Toxic Comment Classification

Fullstack end-to-end toxic comment classification with result interpretation

You can visit the online demo on this page

Data

The dataset is available at kaggle and contains labelled comments from the popular Russian social network ok.ru

Train

python run_training.py

Inference

from src.toxic.inference import Toxic

model = Toxic.from_checkpoint('path_to_model_or_name')
model.infer('привет, придурок')

Result:

{
    'predicted': [
        {'class': 'insult', 'confidence': 0.99324},
        {'class': 'threat', 'confidence': 0.002},
        {'class': 'obscenity', 'confidence': 0.00225}
    ],
    'interpretation': {
        'spans': [(0, 7), (7, 16)],
        'weights': {
            'insult': [-0.34299, 0.93934],
            'threat': [-0.97362, 0.22819],
            'obscenity': [-0.99579, 0.09168]
        }
    }
}

Pretrained model

We provide pretrained model at release page

To download the model execute:

wget https://github.com/esceptico/toxic/releases/download/v0.1.0/model.pth.zip
unzip model.pth.zip

You can download and cache pretrained model by model name as well:

model = Toxic.from_checkpoint('cnn')

List of supported pretrained models:

• cnn: Wide CNN encoder with Feed Forward classification head

Serving

Streamlit

streamlit run ui/app.py -- --model=models/model.pth

Powered By

• captum
• hydra
• tokenizers
• torchmetrics
• streamlit

License

The code is released under MIT license.

The models are distributed under CC BY-NC-SA 4.0 according to the dataset license.