ToChiquinho

ToChiquinho is a toxicity detection system for Brazilian Portuguese texts based on the OLID-BR dataset.

Modeling features

• Early stopping policy (patience).
• Imbalanced dataset handling.
• Hyperparameter optimization (Bayesian optimization).

Usage

ToChiquinho is available as a Docker image. To run it, you need to have Docker installed on your machine.

Then, you can run the following command:

docker run -p 5000:5000 dougtrajano/tochiquinho

Defining the task

ToChiquinho is a toxicity detection system with multiple tasks. You can parameterize the task that will be served by the API by setting the API_TASK environment variable. The available tasks are:

• all: serves all tasks (default) - Warning: this task is not recommended for production environments.
• route: serves the route task. This task is recommended for production environments.
• toxic_spans: serves the toxic spans detection task.
• toxicity_target_type: serves the toxicity target type identification task.
• toxicity_target: serves the toxicity target classification task.
• toxicity_type: serves the toxicity type detection task.
• toxicity: serves the toxicity classification task.

See more in the Environment variables section.

Environment variables

The following environment variables are needed to run the Jupyter Notebooks.

Variable	Description
MLFLOW_TRACKING_URI	The URI of the MLflow tracking server.
MLFLOW_TRACKING_USERNAME	The username of the MLflow tracking server.
MLFLOW_TRACKING_PASSWORD	The password of the MLflow tracking server.
HUGGINGFACE_HUB_USERNAME	The username of the Hugging Face Hub.
HUGGINGFACE_HUB_TOKEN	The token of the Hugging Face Hub.
AWS_PROFILE	The AWS profile to be used.
SAGEMAKER_EXECUTION_ROLE_ARN	The ARN of the SageMaker execution role.

Note: Some variables can be passed to the SageMaker training jobs.

The following environment variables are needed to run the training jobs.

Variable	Description
MLFLOW_TRACKING_URI	The URI of the MLflow tracking server.
MLFLOW_EXPERIMENT_NAME	The name of the MLflow experiment.
MLFLOW_TRACKING_USERNAME	The username of the MLflow tracking server.
MLFLOW_TRACKING_PASSWORD	The password of the MLflow tracking server.
MLFLOW_TAGS	The tags to be added to the MLflow run.
MLFLOW_FLATTEN_PARAMS	Whether to flatten the parameters.
MLFLOW_RUN_ID	The ID of the MLflow run. If set, a child run will be created.
HF_MLFLOW_LOG_ARTIFACTS	Whether to log artifacts to MLflow.
HUGGINGFACE_HUB_TOKEN	The token of the Hugging Face Hub.
SM_TRAINING_ENV	The JSON string of the SageMaker training environment.
TOKENIZERS_PARALLELISM	Whether to use parallelism for tokenizers.

Changelog

See the GitHub Releases page for a history of notable changes to this project.

Contributing

Contributions are welcome! Please open a pull request and include a detailed description of your changes.

Development environment setup

To set up the development environment, you need to have the following tools installed on your machine:

• Docker
• Docker Compose
• Python 3.10

Then, you can run the following commands in the root directory to install the dependencies:

pip install -r requirements-dev.txt
pip install -r requirements-docs.txt
pip install -r requirements.txt

You also need to install the pre-commit hooks:

pre-commit install

Add possible secrets to the .secrets.baseline file

If you get an error when running the pre-commit hooks, you may need to add the possible secrets to the .secrets.baseline file. To do so, you can run the following command:

detect-secrets scan --baseline .secrets.baseline

For more information, see the Yelp/detect-secrets.

Running the tests

To run the tests, you can run the following command in the root directory:

pytest

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.