lizyml 0.1.3

Config-driven ML analysis library for regression and classification

LizyML

LizyML is an analytics library designed to run machine learning workflows in a unified, config-driven way for regression, binary classification, and multiclass classification.

Its goal is to provide a single Model interface that handles the following with stable contracts:

• tune: hyperparameter optimization
• fit: training (CV / refit / early stopping)
• evaluate: evaluation (IF / OOF, before-and-after calibration comparison)
• predict: inference (column mismatch detection, optional explainability)
• export: save and reuse a Model Artifact

Features

• Centralized training control through configuration
• Reproducibility-first design with saved seed / split / params / versions / schema
• Leakage-aware OOF and calibration workflow
• Stable contracts for FitResult / PredictionResult / Artifacts
• Model.load() support for restoring a saved Model Artifact, including training-time evaluation metadata

Current Scope

• LightGBM is the highest-priority backend for the initial phase.
• Interfaces and responsibility boundaries are fixed first so the library can later expand to sklearn and DNNs (Torch).
• Distributed training platforms (Ray / Dask) and large-scale Auto Feature Engineering are not current priorities.

Basic Usage

from lizyml import Model

model = Model(config=config)

model.tune()
fit_result = model.fit()
eval_result = model.evaluate()

pred_result = model.predict(X_test, return_shap=True)

model.export("artifacts/run_001")

Saved Model Artifacts can be restored later with Model.load().

loaded_model = Model.load("artifacts/run_001")

eval_result = loaded_model.evaluate()
pred_result = loaded_model.predict(X_new)

This design lets you inspect not only predictions, but also how well the model performed and which settings were used during training.

Config Example

config = {
    "config_version": 1,
    "task": "regression",
    "data": {
        "path": "data.csv",
        "target": "y",
    },
    "features": {
        "exclude": ["id"],
        "auto_categorical": True,
        "categorical": ["cat_feature1", "cat_feature2"],
    },
    "split": {
        "method": "kfold",
        "n_splits": 5,
        "random_state": 1120,
    },
    "model": {
        "lgbm": {
            "params": {
                "n_estimators": 1000,
                "learning_rate": 0.05,
            }
        }
    },
    "training": {
        "early_stopping": {
            "enabled": True,
            "inner_valid": {
                "method": "holdout",
                "ratio": 0.1,
                "random_state": 1120,
            },
        }
    },
    "tuning": {
        "optuna": {
            "params": {
                "n_trials": 50,
                "direction": "minimize",
            },
            "space": {
                "learning_rate": {
                    "type": "float",
                    "low": 0.01,
                    "high": 0.1,
                    "log": True,
                    "category": "model",
                },
                "num_leaves_ratio": {
                    "type": "float",
                    "low": 0.5,
                    "high": 1.0,
                    "category": "smart",
                },
            },
        }
    },
    "evaluation": {
        "metrics": ["rmse", "mae"],
    },
}

Config Reference (All Keys)

This section documents all config keys currently supported by the implemented schema (config_version=1).

Top-Level Keys

Key	Type	Required	Default	Notes
config_version	int	Yes	-	Only 1 is currently supported.
task	"regression" | "binary" | "multiclass"	Yes	-	Task type used across split/metrics/training behavior.
data	object	Yes	-	Data source and target definition.
features	object	No	{}	Uses exclude=[], auto_categorical=True, categorical=[].
split	object	No (loader fills if missing)	task-dependent	binary/multiclass -> stratified default, regression -> kfold default.
model	object	Yes	-	LightGBM only in current scope.
training	object	No	{}	Uses seed=42 and default early stopping config.
tuning	object | null	No	null	Required only if you call model.tune().
evaluation	object	No	{}	Uses metrics=[] (runtime fallback applies).
calibration	object | null	No	null	Binary-only feature at runtime.

data

Key	Type	Required	Default	Notes
path	str | null	No	null	CSV/Parquet path used when fit() is called without data=.
target	str	Yes	-	Target column name.
time_col	str | null	No	null	Time column for chronological workflows (time_series, purged_time_series, group_time_series require it).
group_col	str | null	No	null	Group column for group-aware split/validation.

features

Key	Type	Required	Default	Notes
exclude	list[str]	No	[]	Columns excluded from training features.
auto_categorical	bool	No	True	Automatically treats suitable columns as categorical.
categorical	list[str]	No	[]	Explicit categorical feature names.

split

split.method is one of:

• kfold
• stratified_kfold
• group_kfold
• time_series
• purged_time_series
• group_time_series

Supported aliases are normalized automatically:

• k-fold -> kfold
• stratified-kfold / stratifiedkfold -> stratified_kfold
• group-kfold / groupkfold -> group_kfold
• time-series / timeseries -> time_series
• purged-time-series / purgedtimeseries -> purged_time_series
• group-time-series / grouptimeseries -> group_time_series

Method-specific keys:

method	Keys
kfold	n_splits=5, random_state=42, shuffle=True
stratified_kfold	n_splits=5, random_state=42
group_kfold	n_splits=5
time_series	n_splits=5, gap=0, train_size_max=null, test_size_max=null
purged_time_series	n_splits=5, purge_gap=0, embargo=0, train_size_max=null, test_size_max=null
group_time_series	n_splits=5, gap=0, train_size_max=null, test_size_max=null

Default when split is omitted:

• task in {"binary", "multiclass"} -> {"method": "stratified_kfold", "n_splits": 5, "random_state": 42}
• task == "regression" -> {"method": "kfold", "n_splits": 5, "random_state": 42, "shuffle": True}

Time-series notes:

• time_series, purged_time_series, and group_time_series all sort rows by data.time_col in ascending order before fold generation.
• train_size_max and test_size_max are shared across all three methods and cap training/validation window sizes.
• purged_time_series uses embargo as the canonical key (embargo_pct is accepted only as a legacy alias during migration).

TimeSeries CV Guide (3 Methods)

All three methods enforce chronological splitting. The difference is how strictly each method blocks potentially leaky rows around the validation window.

Shared index-building rules:

1. Sort rows by data.time_col in ascending order.
2. Build each fold in chronological order (train always before valid).
3. Apply method-specific exclusion (gap / purge_gap / embargo).
4. Apply train_size_max / test_size_max caps when configured.

Quick comparison:

method	boundary key	extra exclusion key	group-safe split	typical use
time_series	gap	-	No	Standard forward CV
purged_time_series	purge_gap	embargo	No	Leakage-sensitive time labels/features
group_time_series	gap	-	Yes	Entity blocks + chronology

1) time_series

Use this when regular forward-chaining CV is enough.

time ---> older ........................................ newer

Fold k:
[           train           ][gap][    valid    ]
                (optional train_size_max / test_size_max caps)

• Validation always comes after training in time.
• gap removes rows right before validation.

2) purged_time_series

Use this when labels/features can leak across nearby timestamps and you need stronger exclusion.

time ---> older ........................................ newer

Fold k:
[      candidate train region      ][purge_gap][ valid ][embargo]
         \____________ train kept _______________/   \__ excluded __/

• purge_gap separates train and validation.
• embargo additionally excludes rows adjacent to the validation window.
• In migration periods, embargo_pct is normalized to embargo.

3) group_time_series

Use this when samples must be split by group blocks while still respecting chronological order.

time-sorted groups:   G1   G1   G2   G2   G3   G3   G4   G4

Fold k:
[       train groups        ][gap][ valid groups ]

• Group boundaries are preserved (no train/valid overlap on the same group).
• Group ordering follows chronological order from time_col.

model

Current backend is LightGBM only.

Accepted input styles:

• BLUEPRINT style: {"model": {"lgbm": {...}}}
• Normalized style: {"model": {"name": "lgbm", ...}}

model.lgbm / normalized model keys:

Key	Type	Required	Default	Notes
name	"lgbm"	Yes (normalized style)	-	Automatically derived from BLUEPRINT style.
params	dict[str, Any]	No	{}	Passed to LightGBM adapter.
auto_num_leaves	bool	No	True	Auto-resolves num_leaves from depth logic.
num_leaves_ratio	float	No	1.0	Must satisfy 0 < ratio <= 1.
min_data_in_leaf_ratio	float | null	No	0.01	Must satisfy 0 < ratio < 1 if set.
min_data_in_bin_ratio	float | null	No	0.01	Must satisfy 0 < ratio < 1 if set.
feature_weights	dict[str, float] | null	No	null	All values must be > 0.
balanced	bool | null	No	null	null=auto (regression→false, binary/multiclass→true). Classification only.

Validation constraints:

• auto_num_leaves=True and params.num_leaves cannot be specified together.
• min_data_in_leaf_ratio and params.min_data_in_leaf cannot be specified together.
• min_data_in_bin_ratio and params.min_data_in_bin cannot be specified together.

Default LightGBM params applied when not overridden in model.params:

Task-specific defaults:

task	objective	metric
regression	huber	["huber", "mae", "mape"]
binary	binary	["auc", "binary_logloss"]
multiclass	multiclass	["auc_mu", "multi_logloss"]

Common defaults:

param	default
boosting	gbdt
n_estimators	1500
learning_rate	0.001
max_depth	5
max_bin	511
feature_fraction	0.7
bagging_fraction	0.7
bagging_freq	10
lambda_l1	0.0
lambda_l2	0.000001
first_metric_only	False
verbose	-1

Runtime-injected default:

• random_state: uses training.seed (default 42)

training

Key	Type	Required	Default	Notes
seed	int	No	42	Global training seed.
early_stopping	object	No	{}	Early stopping behavior.

training.early_stopping:

Key	Type	Required	Default	Notes
enabled	bool	No	True	Disable to skip inner validation strategy.
rounds	int	No	150	Early stopping rounds passed to adapter.
validation_ratio	float | null	No	0.1	Shorthand for inner validation ratio.
inner_valid	object | null	No	null (auto-resolved)	Explicit strategy config.

training.early_stopping.inner_valid.method variants:

method	Keys
holdout	ratio=0.1, stratify=False, random_state=42
group_holdout	ratio=0.1, random_state=42
time_holdout	ratio=0.1

Resolution rules:

• If inner_valid is not explicitly set, method is auto-resolved from split.method:
  • stratified_kfold -> holdout(stratify=True)
  • group_kfold -> group_holdout
  • time_series -> time_holdout
  • purged_time_series -> time_holdout
  • group_time_series -> group_holdout
  • otherwise -> holdout(stratify=False)
• validation_ratio and inner_valid should not be explicitly set together (except round-trip-equivalent holdout dump values).

tuning

tuning is optional. If present:

Key	Type	Required	Default	Notes
optuna.params.n_trials	int	No	50	Number of optimization trials.
optuna.params.direction	"minimize" | "maximize"	No	"minimize"	Optimization direction.
optuna.params.timeout	float | null	No	null	Optional timeout in seconds.
optuna.space	dict[str, Any]	No	{}	Empty dict triggers task-specific default search space.

optuna.space entry format:

"space": {
    "learning_rate": {
        "type": "float",          # "float" | "int" | "categorical"
        "low": 0.0001,            # for float/int
        "high": 0.1,              # for float/int
        "log": True,              # optional for float/int
        "category": "model",      # optional: "model" | "smart" | "training"
    },
    "validation_ratio": {
        "type": "float",
        "low": 0.1,
        "high": 0.3,
        "category": "training",
    },
}

Default search space used when optuna.space = {}:

Param	Type	Range / Choices	Log	Category
objective	categorical	Task-specific (see below)	-	model
n_estimators	int	600 .. 2500	False	model
early_stopping_rounds	int	40 .. 240	False	training
validation_ratio	float	0.1 .. 0.3	False	training
learning_rate	float	0.0001 .. 0.1	True	model
max_depth	int	3 .. 12	False	model
feature_fraction	float	0.5 .. 1.0	False	model
bagging_fraction	float	0.5 .. 1.0	False	model
num_leaves_ratio	float	0.5 .. 1.0	False	smart
min_data_in_leaf_ratio	float	0.01 .. 0.2	False	smart

Task-specific default objective choices:

• regression: ["huber", "fair"]
• binary: ["binary"]
• multiclass: ["multiclass", "multiclassova"]

When using the default space, these fixed params are also applied to every trial:

• auto_num_leaves=True
• first_metric_only=True
• metric is task-specific:
• for regression: ["huber", "mae", "mape"]
• for binary: ["auc", "binary_logloss"]
• for multiclass: ["auc_mu", "multi_logloss"]

evaluation

Key	Type	Required	Default	Notes
metrics	list[str]	No	[]	Metric names validated per task.

If metrics is empty, runtime defaults are:

• regression: ["rmse", "mae"]
• binary: ["logloss", "auc"]
• multiclass: ["logloss", "f1", "accuracy"]

Supported metric names by task:

task	metrics
regression	rmse, mae, r2, rmsle, mape, huber
binary	logloss, auc, auc_pr, f1, accuracy, brier, ece, precision_at_k
multiclass	logloss, f1, accuracy, auc, auc_pr, brier

calibration

Key	Type	Required	Default	Notes
method	"platt" | "isotonic" | "beta"	No	"platt"	All methods are implemented. beta requires scipy.
n_splits	int	No	5	Number of folds for calibration cross-fit.

Runtime notes:

• Calibration is supported only for task="binary".
• method="beta" is supported (install optional dependency: pip install 'lizyml[calibration]').
• Calibration cross-fit splits inherit split.method and its parameters (e.g. gap, purge_gap, embargo, group boundaries). Only the fold count is overridden by calibration.n_splits.

Loader/Override Behavior

• Config source can be dict, .json, .yaml, or .yml.
• Environment-variable overrides use LIZYML__ prefix and __ nesting separators.
• Example: LIZYML__training__seed=999
• Example: LIZYML__model__lgbm__params__learning_rate=0.01

The config system is designed around the following assumptions:

• unified loading from dict / JSON / YAML
• strict validation with pydantic (extra="forbid")
• CLI and environment-variable overrides
• normalization rules for aliases and deprecated keys

Returned Objects and Saved Outputs

FitResult

Training results are expected to follow a fixed schema, including:

• OOF predictions
• IF predictions
• metrics (with a stable structure for raw and calibrated comparisons)
• per-fold models
• training history (learning_curve, best_iteration, and similar metadata)
• split indices
• data fingerprint
• FeaturePipeline state

PredictionResult

Prediction results are expected to include:

• pred
• proba (for binary classification)
• shap_values (when requested)
• used_features
• warnings

Model Artifact

The artifact produced by export is expected to include at least:

• trained models (fold ensemble and/or refit model)
• FeaturePipeline state
• schema (column names / dtypes / categorical handling)
• calibrator
• metrics / history / fit summary
• config_normalized
• versions / format_version

Core Design Priorities

Reproducibility

Given the same config and seed, the library should reproduce the same split behavior and the same primary outputs.

Leakage Prevention

• OOF predictions must come only from models that did not train on the target row.
• Calibration is OOF-only by design, and the calibrator itself is evaluated with cross-fit.
• Time and group constraints must be preserved across outer / inner / calibration splits.

Separation of Responsibilities

• Model acts as a facade that orchestrates components.
• Splitter is limited to returning indices only.
• Plotting uses existing FitResult data and must not rely on inference or recomputation.

Typical Use Cases

• Reproducing training conditions from config
• Managing CV / OOF / early stopping in a consistent way
• Applying binary calibration safely
• Saving trained models and reloading them later with evaluation metadata intact
• Using stable result contracts with predictable shapes and meanings

Future Expansion

• broader sklearn estimator support
• optional DNN (Torch) support
• multiclass calibration
• ranking tasks
• additional export formats such as ONNX / TorchScript / Booster text

Documentation

• Full implementation spec: BLUEPRINT.md
• Proposal and decision history: HISTORY.md

BLUEPRINT.md is the source of truth for implementation rules. This README is a user-facing summary of that specification.