MELITE — Multi-Model Supervised Learning Evaluator
Description
MELITE is a Python package and command-line application for evaluating and comparing supervised classification and regression models on prepared tabular datasets. It separates hyperparameter tuning from model evaluation, preserves the evidence used for selection, and exports the selected model as a reusable artifact for downstream inference.
MELITE operates at the tabular modeling level. Its established input contract
uses numeric feature matrices (X) and targets (y), regardless of how those
features were produced. CSV classification datasets evaluated exclusively with
CatBoost may instead declare categorical columns and use categorical-only or
mixed numeric/categorical predictors. Inputs may originate from fingerprints,
descriptors, dimensionality-reduction methods, clinical variables, experimental
measurements, industrial features, or other prepared representations.
Classification targets with two distinct labels use the binary evaluation contract; targets with three or more distinct labels use the multiclass contract. MELITE determines this from target cardinality and preserves the original class identities through evaluation, export, and inference.
Purpose
MELITE is designed to make supervised-model comparison and selection explicit, reproducible, and auditable. Its workflow separates stages that are often mixed together in machine-learning workflows:
- hyperparameter tuning;
- model evaluation;
- comparison and selection;
- final fitting on all available data;
- model export and inference.
This separation ensures that, within each outer cross-validation split, the data used to evaluate a tuned model are held out from the hyperparameter search that produced it, while preserving the evidence needed to understand how the competing models performed.
Why Use MELITE?
- Controlled evaluation. Hyperparameter tuning is kept separate from the evidence used to compare models.
- Evidence preservation. Aggregate and fold-level evaluation results are retained for every evaluated model, not only for the selected one.
- Explicit selection. Model selection follows a predefined, task-specific criterion based on cross-validation evidence rather than an informal choice after training.
- Domain-agnostic inputs. MELITE works with prepared tabular data without assuming how the features were generated; native categorical predictors are available for CatBoost-only CSV classification.
- Reusable artifacts. After selection, the chosen model can be fitted on all available data and saved as a model artifact for prediction.
- CLI and Python interfaces. MELITE can be used through its command-line workflow and through a focused public Python API.
What MELITE Does
| MELITE does | MELITE does not |
|---|---|
Evaluate supported classification and regression models on prepared tabular X and y. |
Generate domain-specific features or descriptors. |
| Tune supported models within the evaluation design. | Act as a general AutoML framework. |
| Preserve aggregate and fold-level evaluation evidence. | Generate PCA, UMAP, fingerprints, or other feature representations. |
| Select the best active model using the task-specific primary metric. | Process raw domain-specific inputs. |
Fit and export the selected model as a .pkl artifact. |
Perform automatic feature engineering or feature selection. |
| Run inference from exported model artifacts. | Guarantee a stable 1.0 API yet. |
MELITE records model-evaluation evidence and supports reproducible execution; it does not by itself establish external scientific validity or predictive generalizability.
Evaluation Contract
For each registered dataset, MELITE uses the configured task to apply a
task-specific evaluation contract:
- Classification uses repeated stratified outer cross-validation and stratified inner cross-validation. Models are selected by highest mean outer F1-macro; accuracy and ROC-AUC are also reported. Binary ROC-AUC retains its binary definition, while multiclass ROC-AUC uses probability-based One-vs-One macro averaging.
- Regression uses
RepeatedKFoldfor outer evaluation and shuffledKFoldfor inner optimization. Models are selected by lowest mean outer RMSE; MAE and R² are also reported. - Inner hyperparameter optimization is confined to the training portion of each outer split. LinearRegression, PolynomialRegression, and StackingClassifier are evaluated without a separate Optuna search.
- Exact selection ties preserve the configured model order. Aggregate and per-fold outer evidence are retained for every evaluated model. Classification additionally retains held-out precision, recall, F1-score, and support for every original class in every outer fold.
- After selection, tunable models receive a fresh full-data optimization; non-tunable models are fitted directly. The resulting final parameters are persisted in the task-specific results CSV.
melite exportreconstructs the selected model from those parameters, fits it on all available data, and serializes the artifact without another optimization, cross-validation, or selection step.- Smoke mode is intended for execution checks, not final model selection.
Outer cross-validation estimates performance under the configured evaluation design; it does not establish external predictive generalizability. In repeated cross-validation, the same observation can occur in more than one held-out fold across repeats, so those records are not an independent test population.
Installation
Package Users
Install MELITE in a supported Python environment:
python -m pip install melite
Verify the installation:
melite --version
Contributors and Developers
Clone the repository and install in editable mode with development dependencies:
git clone https://github.com/NanoBiostructuresRG/melite.git
cd melite
conda create -n melite_env python=3.11
conda activate melite_env
python -m pip install -e ".[dev]"
To build the documentation locally, install the docs extra as well:
python -m pip install -e ".[dev,docs]"
mkdocs serve
Quick Start
The bundled example creates a ready-to-run synthetic numeric CSV classification dataset and example configuration for a short classification workflow, so you can verify that MELITE is installed and working. Cloning the repository is not required.
melite example
melite run --smoke --config melite_example/config.toml
melite example creates ./melite_example/ in the current directory.
Next steps
- Prepare and use your own datasets.
- Configure MELITE.
- Run a full evaluation.
- Export a model and use the Python prediction API.
Workflow
CLI Workflow
The command-line interface provides the canonical end-to-end MELITE workflow:
- Register one or more prepared tabular datasets in a TOML configuration file.
- Assign each dataset to classification or regression and choose the active models for the configured tasks.
- Run
melite runto generate evaluation evidence and selected results. - Inspect the classification and/or regression artifact family generated for the registered tasks.
- Run
melite exportfor the selected result you want to preserve as a model artifact. - Use the exported
.pklartifact throughmelite.predict()for inference.
Python Workflow
The Python API is intentionally component-oriented. It exposes configuration, dataset loading, evaluation-evidence plotting, artifact-based prediction, and version metadata as public symbols.
MELITE does not expose the full evaluation orchestration as a stable high-level Python workflow API. For reproducible end-to-end execution, use the CLI and a version-controlled TOML configuration.
Supported Classifiers
MELITE supports the following 11 classifier keys. The public names and model families follow the canonical model catalog.
| Key | Public name | Model family | Tunable | Active by default |
|---|---|---|---|---|
svc |
SVC | Individual — support-vector machine | Yes | Yes |
lr |
LogisticRegression | Individual — logistic regression | Yes | No |
gnb |
GaussianNB | Individual — Gaussian naive Bayes | Yes | No |
knn |
KNeighborsClassifier | Individual — nearest neighbors | Yes | No |
dt |
DecisionTreeClassifier | Individual — decision tree | Yes | No |
rf |
RandomForestClassifier | Ensemble — bagging-derived | Yes | Yes |
ada |
AdaBoostClassifier | Ensemble — adaptive boosting | Yes | No |
xgb |
XGBClassifier | Ensemble — gradient boosting | Yes | Yes |
lgbm |
LGBMClassifier | Ensemble — gradient boosting | Yes | No |
catboost |
CatBoostClassifier | Ensemble — gradient boosting | Yes | No |
stack |
StackingClassifier | Ensemble — stacking | No | No |
The ten individual and ensemble classifiers other than stack have
model-specific Optuna search spaces. stack is a fixed, opt-in composite and
is fitted without a separate Optuna search.
The default configuration is:
[classifiers]
active = ["svc", "rf", "xgb"]
The [classifiers].active list accepts any supported subset in evaluation
order, including a single classifier. Classifiers other than svc, rf, and
xgb remain inactive unless explicitly selected.
Every exported classifier artifact supports both predict() and
predict_proba(). SVC remains one classifier family: MELITE preserves direct
SVC class predictions and its decision function while supplying calibrated
probabilities internally for exported inference. The other supported
classifiers provide their estimator-native probability outputs. SVC,
LogisticRegression, and KNeighborsClassifier use scaling pipelines;
GaussianNB, DecisionTreeClassifier, and the standalone bagging and boosting
classifiers operate directly on the supplied numeric features. The opt-in
StackingClassifier combines SVC, RandomForestClassifier, and XGBClassifier as
base estimators with LogisticRegression as the final estimator and uses
probability-based base-estimator outputs.
CatBoost additionally supports native categorical predictors for CSV
classification datasets. Declare those columns with categorical_features and
select only catboost; undeclared predictors must remain numeric. With no
categorical_features declaration, CatBoost retains the same numeric NumPy
input path as the other numeric classifiers.
Binary and multiclass classification use this same classifier inventory. For
multiclass artifacts, classes_ retains the original labels and defines the
column order returned by predict_proba() and melite.predict().
Tunable classifiers use Optuna with its TPE sampler for hyperparameter
optimization. The public n_trials setting explicitly declares the budget per
search, with a normal default of 100 trials. Searches are sequential and seeded
from the canonical RANDOM_STATE, supporting reproducible optimization under
the same configuration and software environment.
Supported Regressors
MELITE supports the following four regressor keys:
| Key | Public name | Estimator pipeline | Parameter optimization |
|---|---|---|---|
linear |
LinearRegression | LinearRegression |
No |
poly |
PolynomialRegression | StandardScaler → PolynomialFeatures(degree=2, include_bias=False) → LinearRegression |
No |
ridge |
Ridge | StandardScaler → Ridge |
Yes — alpha |
lasso |
Lasso | StandardScaler → Lasso(max_iter=10000) |
Yes — alpha |
Ridge and Lasso optimize alpha over the continuous logarithmic range from
1e-4 to 1e2. PolynomialRegression has a fixed degree of two. Before fitting,
MELITE calculates the resulting feature count and rejects an expansion above
10,000 features, reporting both the original and expanded dimensions.
Input Format
Dataset Registry
Datasets are registered under user-defined [datasets.<dataset_id>] entries.
For example:
[datasets.morgan_r2_2048]
path = "data/morgan_r2_2048.npz"
label_path = "raw/labels.npy"
task = "classification"
family = "fingerprints"
method = "Morgan"
variant = "r2_2048"
description = "Morgan fingerprints, radius 2, 2048 bits"
The optional task field accepts exactly classification or regression and
defaults to classification when omitted. Regression requires a
one-dimensional, finite quantitative target and at least one model in
[regressors].active. Classification-only, regression-only, and mixed-task
registries are supported.
Here, family is optional dataset metadata used to describe the feature
representation. It is unrelated to the model selected or evaluated by MELITE.
Each dataset must define path and the label source required by its format:
label_column for CSV datasets or label_path for NPZ datasets.
Optional metadata fields such
as family, method, variant, level, and description are preserved for
traceability and do not trigger dataset-specific execution logic.
The legacy [benchmark] configuration section remains supported for backward
compatibility. New configurations should use the dataset registry.
Numeric CSV datasets omit categorical_features; every predictor must then be
numeric. CatBoost-only categorical or mixed input is configured explicitly:
[datasets.customer_segments]
path = "data/customer_segments.csv"
label_column = "segment"
categorical_features = ["region", "channel"]
[classifiers]
active = ["catboost"]
[catboost]
boosting_type = "Ordered"
categorical_features must be a non-empty list of unique predictor column
names in a CSV classification dataset; it cannot contain label_column and
requires [classifiers].active = ["catboost"]. Declared categorical values
must be non-empty strings, while all remaining predictors must be numeric.
Categorical-only input declares every predictor; mixed input declares only its
categorical predictors. [catboost].boosting_type accepts "Plain" or
"Ordered"; omitting it preserves the "Plain" behavior.
Exported categorical and mixed CatBoost artifacts retain their feature names,
column order, and categorical schema. Inference with those artifacts requires
a pandas DataFrame with exactly the same columns in the same order and
compatible categorical/numeric types; missing, extra, or reordered columns are
rejected. The schema is stored in the artifact, so the original TOML is not
needed for inference. Numeric artifacts retain the existing two-dimensional
NumPy inference contract.
Array Requirements
A registered .npz dataset must contain an X array. MELITE validates that:
Xis two-dimensional;Xis numeric;- the number of rows in
Xmatches the number of labels iny; - an embedded
yarray, when present, matches the configured label vector.
A typical input layout is:
raw/
└── labels.npy
data/
├── morgan_r2_2048.npz
├── rdkit_descriptors.npz
├── PCA85.npz
└── UMAP90.npz
The filenames and feature families are examples only. MELITE does not require PCA, UMAP, fingerprints, descriptors, or any other specific feature-generation method.
Main Outputs
A standard MELITE workflow produces evaluation artifacts and, when requested, a final exported model:
Classification datasets produce results.txt, results.csv,
evaluations.csv, evaluation_folds.csv, optimization_searches.csv,
optimization_provenance.json, evaluation_class_folds.csv,
evaluation_class_summary.csv, classification_metric_provenance.json, and a
dataset-level F1-macro evidence figure.
Regression datasets produce the separate regression_results.txt,
regression_results.csv, regression_evaluations.csv,
regression_evaluation_folds.csv, regression_optimization_searches.csv, and
regression_optimization_provenance.json family. Mixed-task runs produce both
families. Regression currently produces no F1-macro figure.
The classification artifacts have distinct roles:
results.txt— human-readable summary of the selected results.results.csv— selected classifier result for each dataset and the persisted parameter source used bymelite export.evaluations.csv— aggregate evaluation evidence for every active classifier.evaluation_folds.csv— outer-CV evidence for every dataset, classifier, and outer split.evaluation_class_folds.csv— held-out precision, recall, F1-score, and support for every original class in every outer split.evaluation_class_summary.csv— per-class means and population standard deviations over outer folds, including per-fold support summaries.optimization_searches.csv— one row per completed outer or final optimization search.optimization_provenance.json— the effective optimization and evaluation contract for the run.classification_metric_provenance.json— classification metric definitions and ordered original class identities for each executed dataset.figures/evaluation_f1_macro_<dataset>.png— visualization of the outer-CV F1-macro evidence used for classifier selection.Model_<classifier>_<dataset>.pkl— final full-data fitted model created bymelite export.
The corresponding regression CSV files preserve selected results, aggregate
and fold-level RMSE/MAE/R² evidence, completed Ridge/Lasso searches, and the
effective regression evaluation and optimization contract. Exported regression
models use the same Model_<model>_<dataset>.pkl naming pattern.
See the Output Data Contract for the column-level schemas and semantics.
The evaluation figure is generated from already-computed outer-CV evidence. It does not trigger additional fitting, tuning, cross-validation, or selection.
Configuration
MELITE uses TOML configuration files to keep execution choices explicit and reproducible. Configuration controls, among other settings:
- registered datasets and their metadata;
- dataset prediction tasks;
- active classifiers and regressors;
- random state;
- inner and outer cross-validation settings;
- optimization trial budget (
n_trials); - input and output paths.
Use --config to supply a project-specific configuration:
melite run --config my_config.toml
melite export --config my_config.toml --row 0
melite export --config my_config.toml --csv output/regression_results.csv --row 0
A complete configuration template is available at
examples/melite_run.toml.
Smoke mode can be requested independently from the configuration:
melite run --smoke --config my_config.toml
See the full configuration reference in the project documentation.
Development
Running Tests
Run the test suite:
python -m pytest tests -q
Build the documentation in strict mode:
mkdocs build --strict
Build and check the distributions:
python -m build --no-isolation
python -m twine check dist/*
Run the installed-wheel smoke test:
python scripts/smoke_install_wheel.py
Documentation
The full documentation is published at:
https://nanobiostructuresrg.github.io/melite/
Citation
If you use MELITE in your research, please cite it using the metadata in CITATION.cff or the format below:
Contreras-Torres, F. F., & Murrieta, A. C. (2026). MELITE — Multi-Model Supervised Learning Evaluator. Zenodo. https://doi.org/10.5281/zenodo.20382752
License
This project is licensed under the terms of the
GNU Lesser General Public License v3.0 or later.
SPDX identifier: LGPL-3.0-or-later.
Metadata
Release files for melite 0.3.4
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| melite-0.3.4.tar.gz | 163.9 kB | With provenance | Details for melite-0.3.4.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| melite-0.3.4-py3-none-any.whl | Python 3 | none | any | Details for melite-0.3.4-py3-none-any.whl |
Total release size: 262.0 kB
Release files / melite-0.3.4.tar.gz
| Download URL | melite-0.3.4.tar.gz |
|---|---|
| Size | 163.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
d6d49e7711f46464a4e58779404747ca15b3b8233d820842aba8bb6fbcb79b24
|
|
BLAKE2b-256 checksum How to use checksums |
d80e8674df02f37dfbc9f6199de209f26054aada856ac14c60cd642463214183
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 5, 2026.
Transparency logRelease files / melite-0.3.4-py3-none-any.whl
| Download URL | melite-0.3.4-py3-none-any.whl |
|---|---|
| Size | 98.1 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
dcc34f491da774ae65a77a826f3594eada47d11316081acba10931b0204f42c2
|
|
BLAKE2b-256 checksum How to use checksums |
bdc98cc7085438f84f3e297e69a7119669a67a8e53e0f17f91c373694874ca57
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 5, 2026.
Transparency logSecurity for melite 0.3.4
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for melite-0.3.4.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 5, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.