toxicity-detector 0.1.0b1

An LLM-based pipeline to detect toxic speech.

Toxicity Detector

An LLM-based pipeline to detect toxic speech using language models.

Setup

This project uses uv for dependency management.

Prerequisites

• Python 3.12 or higher
• uv package manager

Installation

1. Install uv (if not already installed):

2. Clone the repository:

   git clone https://github.com/debatelab/toxicity-detector.git
   cd toxicity-detector
   

3. Install dependencies:

   uv sync
   

   This will create a virtual environment and install all dependencies specified in pyproject.toml.

4. Install development dependencies (optional):

   uv sync --group dev
   

Environment Variables

Create a .env file in the project root with the following variables:

# API Keys (by the names as specified in the model config files)

# Optional: Custom app config file path
TOXICITY_DETECTOR_APP_CONFIG_FILE=./config/app_config.yaml

Configuration of the Pipeline

#TODO: Add section on configuring the pipeline using YAML files in the config/ directory.

Running the Pipeline

from toxicity_detector import detect_toxicity, PipelineConfig

# Load pipeline configuration from YAML file
pipeline_config = PipelineConfig.from_file('./config/pipeline_config.yaml')

# The text to analyze for toxicity
input_text = 'Peter is dumn.'

# Run toxicity detection
result = detect_toxicity(
    input_text=input_text,  # The text to be analyzed
    user_input_source=None,  # Optional: identifier for the source of the input (e.g., 'chat', 'comment')
    toxicity_type='personalized_toxicity',  # Type of toxicity analysis to perform
    context_info=None,  # Optional: additional context about the conversation or situation
    pipeline_config=pipeline_config,  # Configuration specifying model, paths, and behavior
    serialize_result=True,  # If True, saves the result to disk as YAML
)

# Display the analysis result and toxicity verdict
print(result.answer['contains_toxicity'])

We also provide an example notebook that demonstrates how to run the toxicity detection pipeline with a Hugging Face API key.

Running the Gradio App

The project includes a Gradio web interface for interactive toxicity detection.

Basic Usage

Run the app using uv:

uv run python src/toxicity_detector/app.py

The app will start and be accessible at http://localhost:7860 by default.

Alternative: Using the uv shell

You can also activate the virtual environment and run the app directly:

# Activate the virtual environment
source .venv/bin/activate  # On Linux/Mac
# or
.venv\Scripts\activate  # On Windows

# Run the app
python src/toxicity_detector/app.py

# or (enables live reloading)
gradio src/toxicity_detector/app.py

Developer Mode

To enable developer mode with additional configuration options, update your config/app_config.yaml:

developer_mode: true

Project Structure

toxicity-detector/
├── config/                          # Configuration files
│   ├── app_config.yaml             # App configuration
│   └── default_model_config_*.yaml # Model configurations
├── src/
│   └── toxicity_detector/
│       ├── __init__.py
│       ├── app.py                  # Gradio web interface
│       ├── backend.py              # Core detection logic
│       └── chains.py               # LangChain pipelines
├── logs/                           # Application logs
├── notebooks/                      # Jupyter notebooks for testing
├── pyproject.toml                  # Project dependencies
└── README.md                       # This file

Development

Code Style

The project follows PEP 8 guidelines with a maximum line length of 88 characters.

Run linting checks:

uv run flake8 src/

Running Tests

Run all tests:

uv run pytest

Run tests with verbose output:

uv run pytest -v

Run a specific test file:

uv run pytest tests/test_config.py

Run tests with coverage report:

uv run pytest --cov=src/toxicity_detector

Alternative: Using the activated virtual environment:

# Activate the virtual environment first
source .venv/bin/activate  # On Linux/Mac
# or
.venv\Scripts\activate  # On Windows

# Then run pytest directly
pytest tests/
pytest tests/test_config.py -v

Working with Notebooks

To use Jupyter notebooks for development:

# Install dev dependencies if not already done
uv sync --group dev

# Start Jupyter
uv run jupyter notebook notebooks/

License

See LICENSE file for details.