CodAdapt

Adaptive coded-memory learning for tabular binary classification, multiclass classification and regression, with a scikit-learn-compatible Python API.

from codadapt import CodAdaptClassifier, CodAdaptRegressor

model = CodAdaptClassifier(random_state=42)
model.fit(X_train, y_train)
prediction = model.predict(X_new)
probability = model.predict_proba(X_new)

Features

• Compact train-only numeric/categorical encoding and residual coded memories.
• Binary and multiclass classification, regression, sample weights and missing data.
• Optional class_weight="balanced", computed internally from positive TRAIN weight masses and combined once with sample_weight; the default remains None.
• Clone/Pipeline-compatible estimators and trusted pickle/joblib persistence.
• Source-bound serving export, integrity checking and fitted-only loading.
• Ordered C++17 inference with Python fallback; training and encoding use Python/NumPy.

No new predictive strategy or universal accuracy advantage is claimed. Class balancing can improve recall while worsening logloss. See API, algorithm, limitations and measured performance.

Installation and verified scope

The supplied verified native wheel is for Kaggle Linux x86-64 / CPython 3.13.15 and the exact dependency profile in compatibility. It is tagged cp313-cp313-linux_x86_64, not manylinux. Install a matching prebuilt wheel with pip; no C++ compiler is needed for that wheel. A pure Python wheel is also provided. Distribution manifests distinguish the originally tested artifacts from refreshed documentation/metadata builds.

See installation for local wheel and source commands. Building the extension requires a C++17 compiler and SDK; Linux builds request it by default. CODADAPT_BUILD_NATIVE=0 requests a pure build. Compilation errors are visible, not silently treated as native success. Metadata's Python range is not certification of every Python/OS combination. macOS and unqualified native Windows distributions are excluded from current native-support claims.

Using the native reader

Normal Linux predictions use Python. The automatic guard is unchanged and only opts into native inference on the historical exact Windows CPython 3.12.10 profile when the extension is present. This is not a fresh Windows qualification. Other profiles retain Python automatically; explicit native requests fail visibly on a missing/invalid extension. DLL/ABI/security errors are not hidden.

For the exact verified Kaggle profile, this version-specific advanced context selects the C++ reader without changing estimator state or model files:

from codadapt._inference import inference_backend

with inference_backend("native"):
    prediction = model.predict(X_new)
    probability = model.predict_proba(X_new)

codadapt._inference is a private, version-specific interface, not a new public estimator parameter. Native usage describes diagnostics and restrictions. No automatic-dispatch or serving-contract migration is included in this repository update.

Evidence and examples

The original owner-provided Kaggle receipts record 1,079 PASS, 3 SKIP, 4 XFAIL, zero failures/errors, 200 actual C++ calls and 35 exact array comparisons. The separate pure consumer records 1,029 PASS, 53 SKIP and 4 XFAIL. These results apply to the exact original hashes in the evidence. They do not qualify rebuilt wheels or every Linux host. A receipt/hash is not an independent cryptographic attestation of the remote runner.

The fixed synthetic benchmark has three tasks and three related splits per task; native and Python predictions are identical. Multiclass batch-2048 median latency reductions were 38.20%, 42.80% and 38.28%. Training remains Python; these are not accuracy improvements or a universal performance ranking. See the benchmark and qualification.

Self-contained examples: binary, multiclass, regression and Pipeline/weights/save-load.

Models, tests and release review

Load only trusted artifacts. Serving images must match their producing source contract and numerical profile. Unchanged model arithmetic does not establish arbitrary cross-version pickle or serving compatibility. See persistence.

Run python tools/verify_repository.py for file integrity. Actual-host build, install, native calls, API, persistence and tests use the bounded manual runners in qualification. Workflow files are procedures, not evidence that a platform passed. No publishing workflow is included.

Version remains 2.1.0. This repository is prepared for owner review; copying or publishing is the owner's decision. Multiclass research is separate and does not gate this engineering delivery. Original implementations and previous deliveries remain preserved. The deferred independent full backup is not replaced by the compact release ZIP; it remains required before replacing/deleting the original implementation. Release procedure.

License

CodAdapt is created by Luca Lullo and released under the MIT License.

Author

Created by Luca Lullo.