#!/usr/bin/env python3
"""Generate a runnable scikit-learn baseline script tailored to a CSV's columns.
Reads the header and a sample of rows, sorts columns into numeric / categorical / datetime /
dropped (IDs, near-unique text), and writes a script that:
- splits honestly: time-ordered holdout (--time-col), grouped (--group-col) or stratified random
- scores DummyClassifier/DummyRegressor, a linear model (LogisticRegression / Ridge) and
HistGradientBoosting inside one Pipeline + ColumnTransformer each (no preprocessing leakage)
- prints a comparison table with the metric family for the task
The generated script needs pandas and scikit-learn >= 1.2. This generator itself is stdlib only.
Usage:
python scaffold_baseline.py data.csv --target churned --time-col signup_date > baseline.py
python scaffold_baseline.py data.csv --target price --drop house_id -o baseline.py
python scaffold_baseline.py data.csv --target churned --group-col customer_id --task classification
Exit codes: 0 written, 2 usage or read error.
"""
import argparse
import csv
import math
import os
import sys
from itertools import islice
MISSING = {"", "na", "n/a", "nan", "null", "none", "?"}
def is_num(v):
try:
return math.isfinite(float(v))
except ValueError:
return False
def looks_date(v):
v = v.strip()
return len(v) >= 8 and v[:4].isdigit() and v[4] in "-/" and v[5:7].isdigit()
def classify(header, rows, target, drop, time_col, group_col):
numeric, categorical, dates, dropped = [], [], [], []
n = len(rows)
for c in header:
if c in (target, time_col, group_col) or c in drop:
continue
vals = [r.get(c, "") for r in rows if (r.get(c) or "").strip().lower() not in MISSING]
if not vals:
dropped.append((c, "empty"))
continue
distinct = len(set(vals))
if all(is_num(v) for v in vals):
if distinct == n and all(float(v).is_integer() for v in vals) and c.lower().endswith("id"):
dropped.append((c, "integer ID"))
elif c.lower().endswith(("zip", "zip_code", "postcode", "code")) and distinct > 50:
categorical.append(c) # numeric-looking codes are categories, not quantities
else:
numeric.append(c)
elif sum(looks_date(v) for v in vals) / len(vals) > 0.9:
dates.append(c)
elif distinct > 0.9 * len(vals):
dropped.append((c, "near-unique text (ID or free text)"))
else:
categorical.append(c)
return numeric, categorical, dates, dropped
TEMPLATE = '''"""Baseline models for {target!r} ({task}). Generated by scaffold_baseline.py; edit freely.
Run: python {script_name} (needs pandas, scikit-learn >= 1.2)
"""
import numpy as np
import pandas as pd
from sklearn.compose import ColumnTransformer
from sklearn.dummy import {dummy}
from sklearn.ensemble import {hgb}
from sklearn.impute import SimpleImputer
from sklearn.linear_model import {linear}
from sklearn.metrics import {metric_imports}
from sklearn.model_selection import {split_import}
from sklearn.pipeline import Pipeline, make_pipeline
from sklearn.preprocessing import OneHotEncoder, StandardScaler
DATA = {data_path!r}
TARGET = {target!r}
NUMERIC = {numeric!r}
CATEGORICAL = {categorical!r}
DATES = {dates!r} # expanded into numeric parts below
# Dropped automatically: {dropped}
df = pd.read_csv(DATA)
df = df[df[TARGET].notna()].copy()
for col in DATES:
d = pd.to_datetime(df[col], errors="coerce")
df[col + "_year"], df[col + "_month"], df[col + "_dow"] = d.dt.year, d.dt.month, d.dt.day_of_week
NUMERIC = NUMERIC + [col + "_year", col + "_month", col + "_dow"]
{target_prep}
{split_code}
def preprocess(scale):
num_steps = [SimpleImputer(strategy="median", add_indicator=True)] + ([StandardScaler()] if scale else [])
num = make_pipeline(*num_steps)
cat = make_pipeline(SimpleImputer(strategy="most_frequent"),
OneHotEncoder(handle_unknown="infrequent_if_exist", min_frequency=10, sparse_output=False))
return ColumnTransformer([("num", num, NUMERIC), ("cat", cat, CATEGORICAL)], remainder="drop")
models = {{
"dummy": Pipeline([("prep", preprocess(False)), ("model", {dummy}({dummy_args}))]),
"linear": Pipeline([("prep", preprocess(True)), ("model", {linear}({linear_args}))]),
"hist_gbm": Pipeline([("prep", preprocess(False)), ("model", {hgb}(random_state=0))]),
}}
rows = []
for name, model in models.items():
model.fit(X_train, y_train)
{score_code}
report = pd.DataFrame(rows).set_index("model")
print(report.round(4).to_string())
{leak_check}print("\\nShip only what beats 'dummy' and today's rule by more than the noise between repeated splits.")
'''
CLS = dict(dummy="DummyClassifier", hgb="HistGradientBoostingClassifier", linear="LogisticRegression",
metric_imports="average_precision_score, balanced_accuracy_score, brier_score_loss, log_loss, roc_auc_score",
dummy_args='strategy="prior"', linear_args="max_iter=2000, C=1.0",
leak_check='if report["roc_auc"].max() > 0.97:\n'
' print("\\nWARNING: ROC-AUC > 0.97 on a first baseline usually means target leakage; "\n'
' "run target-leakage-detector before trusting it.")\n',
target_prep='POSITIVE = df[TARGET].value_counts().idxmin() # minority class as positive\n'
'y_all = (df[TARGET] == POSITIVE).astype(int)\n',
score_code=''' p = model.predict_proba(X_test)[:, 1]
rows.append({"model": name, "roc_auc": roc_auc_score(y_test, p), "pr_auc": average_precision_score(y_test, p),
"log_loss": log_loss(y_test, p), "brier": brier_score_loss(y_test, p),
"balanced_acc@prior": balanced_accuracy_score(y_test, p >= y_train.mean())})''')
REG = dict(dummy="DummyRegressor", hgb="HistGradientBoostingRegressor", linear="Ridge",
metric_imports="mean_absolute_error, mean_squared_error, r2_score",
dummy_args='strategy="median"', linear_args="alpha=1.0",
leak_check='if report["r2"].max() > 0.98:\n'
' print("\\nWARNING: R2 > 0.98 on a first baseline usually means target leakage.")\n',
target_prep="y_all = df[TARGET].astype(float)\n",
score_code=''' p = model.predict(X_test)
rows.append({"model": name, "mae": mean_absolute_error(y_test, p),
"rmse": float(np.sqrt(mean_squared_error(y_test, p))), "r2": r2_score(y_test, p),
"wape": float(np.abs(y_test - p).sum() / np.abs(y_test).sum())})''')
def split_code(time_col, group_col, task, test_frac):
if time_col:
return ("split_import", "train_test_split", f'''df = df.sort_values({time_col!r})
y_all = y_all.loc[df.index]
cut = int(len(df) * {1 - test_frac:g}) # latest {test_frac:.0%} of rows = test set
X_train, X_test = df.iloc[:cut], df.iloc[cut:]
y_train, y_test = y_all.iloc[:cut], y_all.iloc[cut:]
''')
if group_col:
return ("split_import", "GroupShuffleSplit", f'''gss = GroupShuffleSplit(n_splits=1, test_size={test_frac:g}, random_state=0)
train_idx, test_idx = next(gss.split(df, y_all, groups=df[{group_col!r}])) # no {group_col} in both sets
X_train, X_test = df.iloc[train_idx], df.iloc[test_idx]
y_train, y_test = y_all.iloc[train_idx], y_all.iloc[test_idx]
''')
strat = "stratify=y_all, " if task == "classification" else ""
return ("split_import", "train_test_split", f'''X_train, X_test, y_train, y_test = train_test_split(
df, y_all, test_size={test_frac:g}, {strat}random_state=0)
''')
def main():
ap = argparse.ArgumentParser(description="Generate a scikit-learn baseline script for a CSV.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv")
ap.add_argument("--target", required=True)
ap.add_argument("--task", choices=["auto", "classification", "regression"], default="auto")
ap.add_argument("--time-col")
ap.add_argument("--group-col")
ap.add_argument("--drop", nargs="*", default=[])
ap.add_argument("--test-frac", type=float, default=0.25)
ap.add_argument("-o", "--output", help="write here instead of stdout")
a = ap.parse_args()
try:
with open(a.csv, newline="", encoding="utf-8-sig") as fh:
reader = csv.DictReader(fh)
header = reader.fieldnames or []
rows = list(islice(reader, 5000))
except OSError as e:
print(f"error: {e}", file=sys.stderr)
return 2
if a.target not in header:
print(f"error: target {a.target!r} not in columns", file=sys.stderr)
return 2
tvals = {(r.get(a.target) or "").strip() for r in rows} - {""}
task = a.task
if task == "auto":
task = "classification" if len(tvals) <= 20 or not all(is_num(v) for v in tvals) else "regression"
if task == "classification" and len(tvals) != 2:
print(f"note: {len(tvals)} classes; the template scores the minority class one-vs-rest. For multiclass, "
"swap metrics for macro-F1 and log_loss on predict_proba.", file=sys.stderr)
numeric, categorical, dates, dropped = classify(header, rows, a.target, set(a.drop), a.time_col, a.group_col)
_, split_name, code = split_code(a.time_col, a.group_col, task, a.test_frac)
parts = dict(CLS if task == "classification" else REG)
script = TEMPLATE.format(target=a.target, task=task, script_name=os.path.basename(a.output or "baseline.py"), data_path=a.csv,
numeric=numeric, categorical=categorical, dates=dates,
dropped=", ".join(f"{c} ({why})" for c, why in dropped) or "none",
split_import=split_name, split_code=code, **parts)
if a.output:
with open(a.output, "w", encoding="utf-8") as f:
f.write(script)
print(f"wrote {a.output}: {len(numeric)} numeric, {len(categorical)} categorical, {len(dates)} date, "
f"{len(dropped)} dropped", file=sys.stderr)
else:
sys.stdout.write(script)
return 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.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
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.