kaggle-new-competition

name	kaggle-new-competition
description	Scaffold a new tabular-multiclass competition in this kit — download/place data, detect id/target/classes/metric, fill src/config.yaml, pick the fold scheme, and freeze the shared CV split. Use once at the very start, before EDA.

Skill: set up a new competition

Run this once when starting a fresh competition. It gets the contract in place so every downstream agent composes.

Steps

1. Get the data into data/. Either place train, test, sample_submission there manually, or (if the Kaggle CLI is configured):

   kaggle competitions download -c <competition-slug> -p data/ && (cd data && unzip -o '*.zip')
   

2. Inspect the shapes and the target. Load train/test/sample_submission, print: dtypes, the id column (matches sample_submission), the target column, the unique classes and their counts, and the prob-column layout of sample_submission (this is the canonical class order your submission must match).

3. Identify the metric. Read the competition's evaluation page. Map it to one of: logloss | accuracy | macro_f1 | qwk (extend src/metrics.py if it's something else). Getting this wrong wastes the whole effort — confirm it literally.

4. Pick the fold scheme. Independent rows → stratified. Multiple rows per entity (a group_col exists and must not split across train/val) → stratified_group. A time column with test in the future → time. Decide from the data description, not a guess.

5. Fill src/config.yaml — competition, id_col, target_col, metric, n_folds, fold_scheme, group_col, seed, and the data_dir.

6. Freeze the split and baseline:

   python src/make_folds.py         # writes artifacts/folds.parquet — the shared CV
   

   Confirm the printed per-fold class balance looks right. Note the class-prior baseline score in EXPERIMENTS_LOG.md as the floor to beat.

7. Hand off to EDA (kaggle-eda agent / prompt 1, Stage 1).

Guardrails

• Do NOT start modelling before folds.parquet exists — the split is the contract.
• Confirm sample_submission column order = your canonical class order; a silent permutation destroys logloss even with a perfect model.