ToxicClassifier

Toxic Classifiers developed for the GATE Cloud. Two models are available:

• kaggle: trained on the Kaggle Toxic Comments Challenge dataset.
• olid: trained on the OLIDv1 dataset from OffensEval 2019 (paper)

We fine-tuned a Roberta-base model using the simpletransformers toolkit.

Requirements

python=3.8 pandas tqdm pytorch simpletransformers

Pre-defined environments

conda

conda env create -f environment.yml

pip

pip install -r requirements.txt

(if the above does not work or if you want to use GPUs, you can try to follow the installation steps of simpletransformers: https://simpletransformers.ai/docs/installation/

Models

1. Download models from the latest release of this repository (currently available kaggle.tar.gz, olid.tar.gz)
2. Decompress file inside models/en/ (which will create models/en/kaggle or models/en/olid respectively)

Basic Usage

python __main__.py -t "This is a test" (should return 0 = non-toxic)

python __main__.py -t "Bastard!" (should return 1 = toxic)

Options

• t: text
• l: language (currently only supports "en")
• c: classifier (currently supports "kaggle" and "olid" -- default="kaggle")
• g: gpu (default=False)

Output

The output is composed by the predicted class and the probabilities of each class.

REST Service

Pre-built Docker images are available for a REST service that accepts text and returns a classification according to the relevant model - see the "packages" section for more details.