#!/usr/bin/env python3
"""Read a gradient-boosting training log and say what the learning curve means.
Accepted inputs (auto-detected):
LightGBM log [100] training's auc: 0.93 valid_1's auc: 0.81
XGBoost log [99] validation_0-logloss:0.31 validation_1-logloss:0.42 (or train-/eval- names)
CatBoost log 99: learn: 0.31 test: 0.42 best: 0.41 (87) total: 1.2s ...
CSV iteration, <train metric>, <valid metric> columns (names containing train / valid)
For the validation curve it reports the best iteration, best value, train/valid gap at the best
iteration, how many rounds ran after it, whether the curve is still improving at the end, and
the overfitting slope after the best. Then recommends: more rounds or a higher learning rate,
a lower learning rate (best found in very few rounds), stronger regularisation (large gap,
rising validation loss), or early stopping patience.
When the training data is the first eval_set entry (logged as validation_0 / valid_0) and a
validation_1 / valid_1 series exists, the better-scoring of the two is treated as the training
curve; --train-name NAME sets it explicitly.
Metric direction is inferred from the name (auc, ndcg, map, accuracy, r2 -> higher is better;
loss, error, rmse, mae, mape, deviance -> lower) and can be forced with --higher-better.
Usage:
python parse_gbm_log.py train.log
python parse_gbm_log.py history.csv --json
lightgbm_script.py 2>&1 | python parse_gbm_log.py -
Exit codes: 0 parsed, 2 nothing parseable or usage error.
"""
import argparse
import csv
import io
import json
import re
import sys
HIGHER = re.compile(r"(auc|ndcg|map|accuracy|acc\b|r2|precision|recall|f1|gini)", re.I)
LGB = re.compile(r"^\[(\d+)\]\s+(.*)$")
LGB_PART = re.compile(r"([\w.]+)'s ([\w@.:-]+):\s*(-?[\d.eE+-]+)")
XGB_PART = re.compile(r"([\w.]+)-([\w@.:-]+):\s*(-?[\d.eE+-]+)")
CB = re.compile(r"^(\d+):\s+learn:\s*(-?[\d.eE+-]+)\s+test:\s*(-?[\d.eE+-]+)")
def parse(text):
"""Return {(dataset, metric): {iteration: value}}."""
series = {}
lines = text.splitlines()
head = lines[0] if lines else ""
if "," in head and re.search(r"iter|round|epoch", head, re.I):
rows = list(csv.DictReader(io.StringIO(text)))
it_col = next(c for c in rows[0] if re.search(r"iter|round|epoch", c, re.I))
for c in rows[0]:
if c == it_col:
continue
ds = "train" if re.search(r"train|learn", c, re.I) else ("valid" if re.search(r"valid|val|test|eval", c, re.I) else c)
metric = re.sub(r"^(train|training|valid|validation|val|test|eval|learn)_?", "", c, flags=re.I) or c
series[(ds, metric)] = {int(float(r[it_col])): float(r[c]) for r in rows if r[c] not in ("", None)}
return series
for ln in lines:
ln = ln.strip()
m = CB.match(ln)
if m:
it = int(m.group(1))
series.setdefault(("train", "loss"), {})[it] = float(m.group(2))
series.setdefault(("valid", "loss"), {})[it] = float(m.group(3))
continue
m = LGB.match(ln)
if not m:
continue
it, rest = int(m.group(1)), m.group(2)
parts = LGB_PART.findall(rest) or XGB_PART.findall(rest)
for ds, metric, val in parts:
series.setdefault((ds, metric), {})[it] = float(val)
return series
def classify(ds):
if re.search(r"train|learn", ds, re.I):
return "train"
return "valid"
def analyse(series, higher_override, train_name=None):
if train_name:
series = {(("train" if d == train_name else d), m): v for (d, m), v in series.items()}
elif not any(classify(d) == "train" for d, _ in series):
# eval_set=[(X_train, y_train), (X_valid, y_valid)] logs the training data as validation_0 / valid_0
names = sorted({d for d, _ in series if re.search(r"_0$", d)})
for d0 in names:
d1 = d0[:-1] + "1"
for (d, m), v in list(series.items()):
if d == d0 and (d1, m) in series:
hb = higher_override if higher_override is not None else bool(HIGHER.search(m))
a, b = sum(v.values()) / len(v), sum(series[(d1, m)].values()) / len(series[(d1, m)])
if (a > b) if hb else (a < b):
series = {(("train" if dd == d0 else dd), mm): vv for (dd, mm), vv in series.items()}
break
valid = {k: v for k, v in series.items() if classify(k[0]) == "valid"}
if not valid:
raise ValueError("no validation series found (need a validation set in the log)")
key = sorted(valid, key=lambda k: (k[0] not in ("valid_1", "validation_1", "eval", "valid", "test"), k))[0]
ds, metric = key
curve = dict(sorted(valid[key].items()))
higher = higher_override if higher_override is not None else bool(HIGHER.search(metric))
its = list(curve)
vals = [curve[i] for i in its]
best_idx = max(range(len(vals)), key=lambda i: vals[i]) if higher else min(range(len(vals)), key=lambda i: vals[i])
best_it, best_val = its[best_idx], vals[best_idx]
last_it, last_val = its[-1], vals[-1]
train = next((v for (d, m), v in series.items() if classify(d) == "train" and m == metric), None)
gap = None
if train and best_it in train:
gap = (train[best_it] - best_val) if higher else (best_val - train[best_it])
tail = vals[-max(3, len(vals) // 10):]
improving = (tail[-1] > tail[0]) if higher else (tail[-1] < tail[0])
after = len(its) - 1 - best_idx
span = (max(vals) - min(vals)) or 1e-12
degradation = ((best_val - last_val) if higher else (last_val - best_val)) / span
advice = []
if best_idx == len(its) - 1 or (improving and len(its) >= 10 and after < max(5, len(its) // 20)):
advice.append("still improving at the last round: raise n_estimators (or the learning rate); you have not "
"reached the best model yet")
if best_it < 100 and len(its) > best_it * 2:
advice.append(f"best at iteration {best_it}: the learning rate is high for this data; lower it (x0.3-0.5) "
"for a smoother optimum and usually a slightly better score")
if degradation > 0.2 and after > 0:
advice.append(f"validation got {degradation:.0%} of its range worse after the best round: early stopping is "
f"essential; patience ~{max(20, min(300, after // 3))} rounds")
if gap is not None:
rel = gap / (abs(best_val) or 1)
if rel > 0.3:
advice.append(f"large train/valid gap at the best round ({gap:.4g}, {rel:.0%} of the valid value): "
"increase min_child_samples/min_child_weight, lower num_leaves/max_depth, raise "
"reg_lambda, use subsample/colsample < 1")
elif rel < 0.02 and not improving:
advice.append("train and valid are close and flat: the model may be under-fitting; allow more capacity "
"(num_leaves / max_depth) or add features")
if not advice:
advice.append("curve looks healthy: best round well inside the run, moderate gap; keep early stopping")
return {"dataset": ds, "metric": metric, "higher_is_better": higher, "iterations": len(its),
"best_iteration": best_it, "best_value": best_val, "last_iteration": last_it, "last_value": last_val,
"rounds_after_best": after, "train_valid_gap_at_best": gap, "still_improving": improving,
"degradation_after_best": round(degradation, 4), "advice": advice}
def main():
ap = argparse.ArgumentParser(description="Interpret a GBM training log / history.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("log", help="log or CSV file, or - for stdin")
ap.add_argument("--higher-better", choices=["yes", "no"], help="override metric direction")
ap.add_argument("--train-name", help="dataset name that holds the training curve (e.g. validation_0)")
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
text = sys.stdin.read() if a.log == "-" else open(a.log, encoding="utf-8").read()
series = parse(text)
if not series:
raise ValueError("no LightGBM / XGBoost / CatBoost / CSV training lines recognised")
hb = None if a.higher_better is None else a.higher_better == "yes"
r = analyse(series, hb, a.train_name)
except (OSError, ValueError, StopIteration, KeyError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
if a.json:
print(json.dumps(r, indent=2))
return 0
gap = "-" if r["train_valid_gap_at_best"] is None else f"{r['train_valid_gap_at_best']:.4g}"
print(f"{r['dataset']} {r['metric']} ({'higher' if r['higher_is_better'] else 'lower'} is better), "
f"{r['iterations']} logged rounds")
print(f"best {r['best_value']:.5g} at iteration {r['best_iteration']}; last {r['last_value']:.5g} at "
f"{r['last_iteration']} ({r['rounds_after_best']} logged rounds after best); train/valid gap at best {gap}")
for adv in r["advice"]:
print(f"- {adv}")
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.