sklearn-tasks-v1

Task dataset for a scikit-learn RL / eval environment, in the shape used by the Prime Intellect Environments Hub.

25 scikit-learn tasks across 5 categories. Each task names a fixed, seed-generated dataset and an instruction; the answer is the array left in result, compared to a reference. Grading is deterministic — no LLM judge, no external API, no download.

The skill being tested is not array algebra. It is knowing which estimator, which preprocessing and — mostly — which evaluation is correct: fitting a scaler before the split, reporting accuracy on a 90/10 imbalanced problem, or letting a pipeline leak through cross-validation are the mistakes this domain actually punishes, and the tasks are chosen to expose them.

Category	Tasks	Covers
preprocessing	6	StandardScaler, MinMaxScaler, PCA explained variance, SelectKBest, stratified train/test split shapes
classification	6	LogisticRegression, RandomForest, DecisionTree importances, an SVC pipeline, GradientBoosting, raw predictions
validation	5	KFold cross-validation, a scaler fitted inside each fold, score mean/std, KMeans inertia and cluster sizes
metrics	4	accuracy vs precision/recall/F1 on an imbalanced set, ROC AUC from predict_proba, confusion-matrix counts, class balance
regression	4	R²/MAE/RMSE, coefficients, Ridge, a scaled Ridge pipeline

Datasets

All three are generated in-process with a fixed seed — nothing is downloaded:

X_clf, y_clf = make_classification(n_samples=300, n_features=8, n_informative=5,
                                   n_redundant=1, n_classes=2, random_state=20260806)
X_imb, y_imb = make_classification(n_samples=400, n_features=6, n_informative=4,
                                   n_redundant=0, n_classes=2, weights=[0.9, 0.1],
                                   random_state=20260806)
X_reg, y_reg = make_regression(n_samples=200, n_features=5, n_informative=3,
                               noise=8.0, random_state=20260806)

What makes ML gradeable at all

• Every dataset comes from a fixed seed — no download, no network.
• Every estimator with a random_state gets one, and the prompt states it. An unseeded RandomForest is not a gradeable answer.
• References return small arrays of numbers (scores, coefficients, counts), rounded where rounding is safe — never a fitted model object.
• n_jobs=1 everywhere. Thread scheduling changes floating-point reduction order, and an answer key that depends on the core count of the machine that built it is not an answer key.

Verification

Every task is independently checked: it runs, is deterministic across two fresh executions (which is what would catch an unseeded estimator or a thread-order dependence), returns finite JSON-safe numbers, is non-empty, and survives the serialisation round-trip exactly. All 25 pass on scikit-learn 1.9.0.

Builder and verifier: build_tasks.py.