IBM Breast Cancer Workshop

Human-in-the-loop ML competition on Wisconsin-style cytology features: maximize ROC-AUC on cross-validated data/public.csv, then fit on full public data, evaluate on data/test/, and push reports to Skore Hub.

Where this repo stands

Iteration 1 is done. A uniform random classifier baseline (01_baseline) is implemented, evaluated (ROC-AUC on an 80/20 hold-out), and recorded in journal/JOURNAL.md. EDA lives under data/eda/eda.md.

Check the Skore Hub

Submissions land in workspace ibm-workshop/competition at https://skore.probabl.ai. Each report key is {SKORE_USERNAME}/{experiment_stem} (set SKORE_USERNAME in .env).

Agent authentication uses SKORE_HUB_API_KEY from .env — no interactive login.

Before you start — do this once

1. Set up Python (uv)

# install uv if needed: curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync --all-groups
set -a && source .env && set +a
uv run python -c "import skore, skrub, ibm_workshop; print('imports OK')"

2. Configure Skore Hub

Edit .env and set:

• SKORE_USERNAME — your username for the leaderboard

3. Sanity check

uv run pytest tests/smoke/test_01_baseline.py -q