Weak Supervision Demo

A lightweight Python project that demonstrates how multiple labeling functions can generate training labels without manually annotating every example.

Features

• Rule-based labeling functions
• Positive, negative, and abstain labels
• Majority-vote label aggregation
• Coverage and label-distribution metrics
• JSON report generation
• Automated Pytest test suite
• Docker support
• GitHub Actions continuous integration

How It Works

The project applies several independent labeling functions to each text example.

Each labeling function returns one of these values:

• POSITIVE
• NEGATIVE
• ABSTAIN

The votes are combined using majority voting. A tie produces an abstain result.

Project Structure

.
├── .github/
│   └── workflows/
│       └── ci.yml
├── tests/
│   └── test_labeling.py
├── .gitignore
├── Dockerfile
├── labeling.py
├── README.md
├── requirements.txt
└── train.py

Installation

Create a virtual environment:

python -m venv venv

Activate it:

.\venv\Scripts\Activate.ps1

Install dependencies:

python -m pip install -r requirements.txt

Run Tests

python -m pytest -q

Expected result:

11 passed

Run the Demo

python train.py

This creates:

weak_supervision_report.json

Use a Custom Text File

Create a file with one example per line:

I love this product.
This is terrible.
The package arrived today.

Run:

python train.py --input examples.txt

Use a JSON File

Example:

[
  {"text": "This product is amazing."},
  {"text": "I want a refund."}
]

Run:

python train.py --input examples.json

Docker

Build the image:

docker build -t weak-supervision-demo .

Run it:

docker run --rm weak-supervision-demo

Technology

• Python 3.11
• Weak supervision
• Rule-based labeling
• Pytest
• Docker
• GitHub Actions

License

MIT