Kaggle Starter Toolkit (Google Tunix Hackathon)

This repo is a lightweight, competition-agnostic Kaggle starter kit. It includes a baseline training script, a prediction script for generating submissions, and a simple data pipeline that handles numeric + categorical columns.

Quickstart

1) Install dependencies

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2) Download data

Use the Kaggle CLI (install separately with pip install kaggle and set up your Kaggle API token) to download the competition data into data/:

mkdir -p data
kaggle competitions download -c google-tunix-hackathon -p data
unzip -o data/google-tunix-hackathon.zip -d data

If the download name differs, list files with ls data and adjust the paths below.

3) Train a baseline model

Set the target column (and optionally ID column) based on the competition data schema. The ID column is excluded from training features automatically.

export TARGET_COLUMN=target
export ID_COLUMN=id
python -m src.train --train-path data/train.csv

This will create artifacts/model.joblib and print a local validation score. You can also add --cv 5 to report a cross-validation score or --model-type random_forest for a stronger baseline.

4) Generate a submission

python -m src.predict \
  --test-path data/test.csv \
  --model-path artifacts/model.joblib \
  --output-path submissions/submission.csv

If the competition expects probabilities instead of hard class labels, add --use-proba to the prediction command. For multi-class classification, this will emit one column per class (e.g. target_classA, target_classB).

Configuration

Environment variables:

• TARGET_COLUMN (required): name of the target column in train.csv.
• ID_COLUMN (optional): name of the ID column used for submissions.
• PROBLEM_TYPE (optional): classification or regression. If omitted, the script infers it.

Project Layout

.
├── data/                  # Kaggle data (not tracked)
├── src/
│   ├── data.py             # I/O and schema utilities
│   ├── model.py            # Model + preprocessing pipeline
│   ├── train.py            # Training entry point
│   └── predict.py          # Submission generation
├── artifacts/              # Trained models (not tracked)
└── submissions/            # Submission files (not tracked)

Notes

This is intentionally minimal. Once you know the metric and data quirks, you can customize:

• Feature engineering in src/model.py
• Validation logic in src/train.py
• Custom metrics in src/train.py