TextDetoxification

This repository contains a Text Detoxification project - transforming the text with toxic style into the text with the same meaning but with neutral style. I have tried different models to solve this problem, the pre-trained T5 shows the best performance. You can download model here

Installation

1. Download repository

git clone https://github.com/slewie/TextDetoxification

2. Install libraries

pip install -r requirements.txt

3. Download and prepare dataset:

python ./src/data/make_dataset.py

Train model

python src/models/train_model.py

with such parameters:

• --model_name | what is model will be trained. Now supported only ['toxicity_identifier'] from pytorch and all models with tokenizer from transformers

• --library | parameter that corresponds to which library the model is from: pytorch or transformers

• --num_epochs | number of model training epochs

• --random_seed | parameter responsible for reproducible results

• --device | cuda or cpu

• --learning_rate | optimizer learning rate

• --data_path | path to the .csv file with data

• --vocab_size | vocabulary size for the model

• --save_model | save model or not

Predict model

python src/models/predict_model.py

with such parameters:

• --model_name | what is model will be trained.
• --library | parameter that corresponds to which library the model is from: pytorch or transformers

After running the command, you need to write a text sequence that needs to be predicted or generated.