Deep Learning Mini-Project (EE-559) — Group 8

DetoxText: Text Detoxification Using Finetuned Encoder-Decoder Models

Project Overview

This repository contains the implementation of our Deep Learning Mini-Project for EE-559 at EPFL. The project focuses on fostering safer online spaces by developing deep learning models that detect hate speech in various formats, including text, images, memes, videos, and audio content.

Objectives

• Develop deep learning models that accurately classify hate speech while minimizing false positives.
• Evaluate model performance with benchmarks and interpretability metrics.
• Address ethical and legal considerations in AI-powered content moderation.

Project Structure

The repository contains the following directories and files:

Directories

• cleaned_data/ contains the preprocessed data
• data/ contains the raw data that we download (paradetox.tsv)
• data_preprocessing/ contains scripts for loading and preprocessing datasets.
• eval/ includes evaluation routines and performance metrics.
• trainer/ implements the training loop and model management.
• utils/ provides utility functions used across the project.
• visualizatios/ contains a folder with the plots used in our report.

Top-Level Files

• README.md – Main project documentation.
• requirements.txt – Python dependencies.
• main_config.yaml – Configuration file for training and evaluation.
• main.py – Entry point to run the training pipeline.
• basic_running_scripts.sh – Shell script(s) to launch experiments.
• tokens.yaml – Contains token/API configuration if needed.
• starter.sh - Center script that runs the training

Getting Started

1. Clone the repository:

   git clone https://github.com/charafkamel/Deep_Learning.git
   cd Deep_Learning

Setup & Installation

2. Create a virtual environment (optional)

   python -m venv venv
   source venv/bin/activate  # On Windows use `venv\Scripts\activate`

3. Install dependencies

   pip install -r requirements.txt

Dataset

To train, evaluate and test the model we use the paradetox dataset from https://github.com/s-nlp/paradetox/blob/main/paradetox/paradetox.tsv. It is a parallel detoxification dataset containing over more than 12,000 toxic-neutralized sentence pairs. Collected via crowdsourcing using Reddit, Twitter and Jigsaw. Ensures semantic similarity and fluency while removing toxicity.

Model Architecture

We use the Qwen3-0.6B model and the T5-base model as our primary architectures, representing decoder-only and encoder-decoder architectures respectively.

Training

This project supports Supervised Fine-Tuning (SFT) and Reinforcement Learning (GRPO) training for both encoder-decoder (Seq2Seq) and decoder-only models. Below are the available training configurations and how to use them.

Available Training Setups

We support five different training setups:

1. Encoder-Decoder (Seq2Seq) Models

• --base: Train using standard cross-entropy loss
• --count: Train using our custom loss function

2. Decoder-Only Models

• --base_generative: Train using standard cross-entropy loss
• --count_generative: Train using our custom loss function
• --rl: Train using Reinforcement Learning (GRPO) with our custom reward function

How to Train

Run the following command with the appropriate flag:

python main.py --base              # Seq2Seq with standard loss
python main.py --count             # Seq2Seq with custom loss
python main.py --base_generative   # Decoder-only with standard loss
python main.py --count_generative  # Decoder-only with custom loss
python main.py --rl                # Decoder-only with RL (GRPO)

Evaluation

To evaluate model performance:

python eval/eval.py

This will run the evaluation script on all of the saved models under the defined hugging-face account (defined in main_config.yml with the "hf_username" parameter). The results will be saved in a csv under

Run everything

It is also possible to run all of the training using the starter script:

./starter.sh

This will launch all 5 training jobs in parallel. After those are finished running, we might launch the evaluation script

Contributors

• Kamel Charaf
• Efe Tarhan
• Mahmut Serkan Kopuzlu