#!/usr/bin/env python3
"""Check a CatBoost parameter set (JSON) for conflicts and risky choices before a long training run.
Rules (errors were confirmed to raise CatBoostError in CatBoost 1.2):
CB001 error more than one of iterations / n_estimators / num_boost_round / num_trees
CB002 error both learning_rate and eta
CB003 error more than one of scale_pos_weight / class_weights / auto_class_weights
CB004 error bootstrap_type="Bayesian" with subsample; bagging_temperature without Bayesian bootstrap
CB005 error depth > 16
CB006 error use_best_model=True without an eval set (pass --has-eval-set when you will provide one)
CB007 warning max_leaves set but grow_policy is not "Lossguide" (silently ignored)
CB008 warning no overfitting detector (od_type / early_stopping_rounds) with many iterations
CB009 warning loss_function / eval_metric do not match the task (--task)
CB010 warning depth >= 10 (symmetric trees: 2^depth leaves; slow, overfits)
CB011 warning learning_rate >= 0.3 with >= 1000 iterations, or <= 0.005 (very slow)
CB012 info task_type="GPU" with rsm (feature subsampling) on non-pairwise losses is unsupported on GPU
CB013 info one_hot_max_size > 255 or large: one-hot for many levels loses CatBoost's CTR advantage
CB014 warning l2_leaf_reg <= 0
Usage:
python catboost_params_check.py params.json --task binary
python catboost_params_check.py params.json --task regression --has-eval-set --json
echo '{"iterations": 5000, "n_estimators": 100}' | python catboost_params_check.py -
Exit codes: 0 no errors, 1 errors, 2 read error.
"""
import argparse
import json
import sys
TASK_LOSSES = {
"binary": {"Logloss", "CrossEntropy", "Focal"},
"multiclass": {"MultiClass", "MultiClassOneVsAll"},
"regression": {"RMSE", "MAE", "Quantile", "Expectile", "Huber", "Poisson", "Tweedie", "MAPE", "LogLinQuantile",
"Lq", "RMSEWithUncertainty", "LogCosh", "Cox", "SurvivalAft"},
"ranking": {"YetiRank", "YetiRankPairwise", "PairLogit", "PairLogitPairwise", "QueryRMSE", "QuerySoftMax",
"LambdaMart", "StochasticRank"},
}
TASK_METRICS = {
"binary": {"AUC", "PRAUC", "Logloss", "CrossEntropy", "F1", "Precision", "Recall", "Accuracy", "BalancedAccuracy",
"MCC", "BrierScore", "Kappa", "NormalizedGini"},
"multiclass": {"MultiClass", "Accuracy", "TotalF1", "AUC", "MCC", "Kappa", "WKappa", "HammingLoss"},
"regression": {"RMSE", "MAE", "R2", "MAPE", "SMAPE", "Quantile", "Poisson", "Tweedie", "Huber", "MSLE", "MedianAbsoluteError",
"Expectile", "LogCosh"},
"ranking": {"NDCG", "PFound", "MAP", "MRR", "ERR", "QueryRMSE", "PrecisionAt", "RecallAt", "AUC"},
}
def base(name):
return str(name).split(":")[0] if name is not None else None
def check(p, task, has_eval):
out = []
def add(code, sev, msg):
out.append({"code": code, "severity": sev, "message": msg})
iters = [k for k in ("iterations", "n_estimators", "num_boost_round", "num_trees") if k in p]
if len(iters) > 1:
add("CB001", "error", f"only one of {iters} may be set")
if "learning_rate" in p and "eta" in p:
add("CB002", "error", "set learning_rate or eta, not both")
w = [k for k in ("scale_pos_weight", "class_weights", "auto_class_weights") if k in p]
if len(w) > 1:
add("CB003", "error", f"only one of {w} may be set")
bt = p.get("bootstrap_type")
if bt == "Bayesian" and "subsample" in p:
add("CB004", "error", "Bayesian bootstrap does not support subsample (use bagging_temperature)")
if "bagging_temperature" in p and bt not in (None, "Bayesian"):
add("CB004", "error", f"bagging_temperature needs bootstrap_type='Bayesian' (got {bt!r})")
depth = p.get("depth", p.get("max_depth"))
if isinstance(depth, (int, float)):
if depth > 16:
add("CB005", "error", f"depth {depth} > 16 (CatBoost maximum)")
elif depth >= 10:
add("CB010", "warning", f"depth {depth}: symmetric trees have 2^depth leaves; try 4-8 first")
if p.get("use_best_model") and not has_eval:
add("CB006", "error", "use_best_model=True requires eval_set in fit() (or pass --has-eval-set)")
if "max_leaves" in p and p.get("grow_policy") != "Lossguide":
add("CB007", "warning", "max_leaves only applies with grow_policy='Lossguide'; it is ignored otherwise")
n_it = next((p[k] for k in iters), 1000) if iters else 1000
if not any(k in p for k in ("od_type", "od_wait", "early_stopping_rounds")) and n_it >= 500:
add("CB008", "warning", f"{n_it} iterations without an overfitting detector: add od_type='Iter', od_wait=200 "
"(or early_stopping_rounds) with an eval_set")
loss, metric = base(p.get("loss_function")), base(p.get("eval_metric"))
if task:
if loss and loss not in TASK_LOSSES[task]:
add("CB009", "warning", f"loss_function {p['loss_function']!r} is unusual for a {task} task "
f"(expected one of {sorted(TASK_LOSSES[task])[:6]}...)")
if metric and metric not in TASK_METRICS[task]:
add("CB009", "warning", f"eval_metric {p['eval_metric']!r} is unusual for a {task} task")
lr = p.get("learning_rate", p.get("eta"))
if isinstance(lr, (int, float)):
if lr >= 0.3 and n_it >= 1000:
add("CB011", "warning", f"learning_rate {lr} with {n_it} iterations: expect early overfitting; use 0.03-0.1")
if lr <= 0.005:
add("CB011", "warning", f"learning_rate {lr} is very low: training will need many thousands of iterations")
if p.get("task_type") == "GPU" and "rsm" in p and loss not in ("PairLogitPairwise", "YetiRankPairwise"):
add("CB012", "info", "rsm (feature subsampling) on GPU is only supported for pairwise losses")
ohs = p.get("one_hot_max_size")
if isinstance(ohs, int) and ohs > 255:
add("CB013", "info", f"one_hot_max_size {ohs}: one-hot encoding many levels gives up CatBoost's ordered CTRs")
l2 = p.get("l2_leaf_reg")
if isinstance(l2, (int, float)) and l2 <= 0:
add("CB014", "warning", "l2_leaf_reg <= 0 removes leaf regularisation; typical range 1-10")
return out
def main():
ap = argparse.ArgumentParser(description="Check CatBoost parameters for conflicts.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("params", help="JSON object of CatBoost parameters, or - for stdin")
ap.add_argument("--task", choices=["binary", "multiclass", "regression", "ranking"])
ap.add_argument("--has-eval-set", action="store_true", help="you will pass eval_set to fit()")
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
p = json.loads(sys.stdin.read() if a.params == "-" else open(a.params, encoding="utf-8").read())
if not isinstance(p, dict):
raise ValueError("parameters must be a JSON object")
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
findings = check(p, a.task, a.has_eval_set)
errors = sum(f["severity"] == "error" for f in findings)
if a.json:
print(json.dumps({"errors": errors, "findings": findings}, indent=2))
else:
for f in findings:
print(f"{f['code']} [{f['severity']}] {f['message']}")
print(f"\n{len(findings)} finding(s), {errors} error(s)" + ("" if findings else ": parameters look consistent"))
return 1 if errors else 0
if __name__ == "__main__":
sys.exit(main())
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.
counterfactual-explanation-generator
Generate counterfactual explanations and algorithmic recourse: the smallest realistic change to a row's features that changes the model's decision. Solves exact minimal-cost recourse for linear and logistic models (MAD-scaled L1 cost, immutable features, increase/decrease directions, bounds, integer and binary features, diverse alternatives, features every counterfactual must change, and a no-recourse finding when immutable features decide), generates DiCE counterfactuals for any scikit-learn model with constraints turned into permitted ranges and every result re-predicted, and validates counterfactuals from any method for constraint violations, implausible values, off-manifold rows, broken derived-feature relations and failure to flip the prediction.