Based on Mitchell et al. (2019), "Model Cards for Model Reporting", the Hugging Face model card guide, and common model-risk review practice (e.g. SR 11-7 style documentation).
Section by section
Model details. One paragraph that a non-specialist can follow: what is predicted, for whom, when it runs, and what happens with the score. Add the type, version, date, owners, licence and framework.
Bad: "A gradient boosting model for churn." Good: "Scores active consumer subscribers every Monday with the probability of cancelling within 60 days. The top-scored customers are offered a retention call by the retention team."
Intended use / out of scope. Name the decision and the users. Out-of-scope entries stop misuse: other populations, other decisions, automated use without review, uses with legal effects.
Data. Give the source, the period, the size, the label definition (exactly when an event counts), the sampling and filtering, and known gaps (segments absent or under-represented). Evaluation data must be held out from training. Say whether it is random, grouped or out-of-time.
Evaluation. Use the metrics the decision needs (see ml-success-metric-selector), with 95% CIs, on the evaluation data. Add the baseline (the current process or a simple rule), the decision threshold and its rationale, and the calibration if probabilities are used directly.
Performance across groups. Use the same metrics for each slice that matters: protected or sensitive groups (where lawful to use for evaluation), customer segments, regions, channels, and data-quality strata (new vs long-standing customers). Mark small slices. Explain what is done about gaps.
Fairness. State which attributes were assessed, which fairness notion fits the decision and why, the measured gaps, and the decision taken (accept with reason, mitigate, restrict use).
Explainability. Global drivers, and how individual decisions are explained (reason codes, SHAP background, LIME), with links to the reports.
Limitations. List concrete conditions where the model is unreliable: thin-file customers, new products, regime changes, extrapolation ranges, known label noise.
Ethical considerations. Cover who is affected and how, sensitive features and their proxies, potential harms (denial of service, unequal treatment, privacy), and the human oversight in place.
Risks and mitigations. Each risk gets a mitigation and an owner. Unmitigated risks are stated as accepted.
Monitoring and maintenance. List the monitored metrics (with label delay), drift checks, alert thresholds, the retraining cadence or trigger, the owner, and the decommission criteria.
Words to avoid
| Avoid | Write instead |
|---|---|
| "unbiased", "fair", "bias-free" | the measured gaps and the notion used |
| "state of the art" | the comparison: "+0.11 AUC over the current rule" |
| "highly accurate" | the metric, value and CI |
| "works for all customers" | the populations evaluated and the slices with low n |
| "will be monitored" | what, how often, the threshold, and who |
Review checklist
- a non-specialist can tell what the model does and must not be used for
- data periods and label definition stated; evaluation data held out (out-of-time if applicable)
- every metric has a CI and a data split; baseline and threshold given
- performance by group with CIs; small slices marked; gaps addressed
- limitations and ethical considerations specific to this model
- risks have mitigations and owners
- monitoring plan with thresholds and owner
- version, date and contact; no TODOs (
card_lint.pyclean)
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.