Adapt Automl¶

adapt_automl — mvp.adapt_automl

Cluster: ML & Optimisation | Type: component | MCP Tools: 46

Overview¶

Comprehensive AutoML system with FLAML, AutoGluon, H2O, and sklearn backends. Accepts dict records (from adapt_pandas) or raw float arrays, handles automatic preprocessing (imputation, encoding, scaling), and persists trained models with a 31-tool MCP sub-package covering training, evaluation, model versioning, ensembling, explainability, and an advisory adapter guide.

When to use:

• Automatically selecting and tuning ML models for classification or regression tasks
• Integrating tabular AutoML into an agent pipeline via MCP tools
• Benchmarking multiple backends (FLAML, AutoGluon, H2O) with a unified interface

Example:

from mvp.adapt_automl import AdaptAutoMLBlock, AutoMLInput

block = AdaptAutoMLBlock(name="automl")
result = block.infer(AutoMLInput(
    records=[{"age": 25, "income": 50000}, {"age": 40, "income": 80000}],
    y=[0, 1],
    task="classification",
    time_budget_sec=60,
))
# result.ok → True; result.value → AutoMLOutput with best_model, best_score, leaderboard

Works well with: adapt_pandas, adapt_sklearn, align_evals

Validation Notes¶

AutoML scores are strongest when they come from cross-validation or an explicit holdout set. For very small or heavily imbalanced classification datasets, the sklearn fallback may be unable to run reliable cross-validation because one class has too few examples. In that case leaderboard entries are marked with validation: "training_fallback" and ranked by training-set score so the workflow can still complete.

Treat training_fallback scores as a smoke-test signal only. Add more labelled examples or evaluate the saved model with automl_fit_eval, automl_evaluate, or automl_compare_holdout before using the model for production decisions.

Public API¶

PandasToAutoMLAdapter(AIBlock[DataOutput, AutoMLInput, None])¶

Bridge adapt_pandas DataOutput to adapt_automl AutoMLInput.

Methods:

infer(data: DataOutput) -> Result[AutoMLInput]¶

AutoMLGuideRecommendation¶

Validated adapter recommendation decision record.

Methods:

to_dict() -> dict[str, Any]¶

The advisory payload as adapter_guide.recommend_adapter returns it.

to_metadata() -> dict[str, Any]¶

AutoMLDecisionError(ValueError)¶

The LLM did not produce a usable, validated adapter recommendation.

LLMAutoMLRuntime¶

Provider-neutral adapter-recommendation runtime over G6's LLM caller.

Constructor:

Methods:

recommend(inputs: dict[str, Any], floor: 'AutoMLGuideRecommendation', context: str = '') -> Result[dict]¶

AutoMLGuidePlanner¶

Runtime-first facade with the deterministic floor as honest fallback.

Constructor:

Methods:

recommend(data_type: str = 'tabular', data_size: str = 'medium', goal: str = 'predict', expertise: str = 'beginner', notes: str = '') -> AutoMLGuideRecommendation¶

AdaptAutoMLBlock(AIBlock[AutoMLInput, AutoMLOutput, None])¶

Comprehensive AutoML block with multi-backend support and preprocessing.

Methods:

infer(data: Any) -> Result[AutoMLOutput]¶

PreprocessingConfig(BaseModel)¶

Controls automatic preprocessing. All fields optional with sensible defaults.

AutoMLInput(BaseModel)¶

Unified input for AutoML — accepts dict records or float arrays.

LeaderboardEntry(BaseModel)¶

A candidate model on the leaderboard.

FeatureImportance(BaseModel)¶

Feature importance score.

PreprocessingReport(BaseModel)¶

Report of preprocessing steps applied.

AutoMLOutput(BaseModel)¶

Output from the AutoML system.

AdaptAutoMLMCPBlock(AIBlock[MCPAutoMLInput, MCPAutoMLOutput, dict])¶

MCP AutoML block with persistence and model caching.

Methods:

infer(inp: MCPAutoMLInput) -> Result[MCPAutoMLOutput]¶

MCPAutoMLInput(BaseModel)¶

Input for all 31 MCP operations, discriminated by op field.

MCPAutoMLOutput(BaseModel)¶

Output from all 31 MCP operations.

AutoMLStore¶

SQLite store for AutoML MCP persistence.

Constructor:

Methods:

record_readiness(model_name: str, record: dict[str, Any]) -> None¶

Store the readiness snapshot for model_name.

get_readiness(model_name: str) -> dict[str, Any] | None¶

schema_version() -> int¶

upsert_model(name: str, backend: str, task: str, metric: str, score: float, n_models_tried: int, model_path: str, preprocessing_json: str, feature_columns_json: str) -> None¶

get_model(name: str, version: int | None = None) -> dict[str, Any] | None¶

list_models() -> list[dict[str, Any]]¶

list_model_versions(name: str) -> list[dict[str, Any]]¶

delete_model(name: str) -> bool¶

save_model_blob(name: str, blob: bytes) -> None¶

get_model_blob(name: str) -> bytes | None¶

add_experiment(model_name: str, backend: str, task: str, metric: str, score: float, n_models_tried: int, notes: str, test_score: float = 0.0, test_size: float = 0.0) -> None¶

list_experiments(backend: str = '', task: str = '') -> list[dict[str, Any]]¶

add_prediction(model_name: str, n_samples: int, backend: str) -> None¶

register_dataset(name: str, n_samples: int, n_features: int, task: str, notes: str) -> None¶

get_dataset(name: str) -> dict[str, Any] | None¶

count_all() -> dict[str, int]¶

Functions¶

agentic_planner_enabled(default_enabled: bool = True) -> bool¶

Decide whether the agentic adapter-planner should be used.

validate_guide_recommendation(rec: Any, allowed: set[str] | frozenset[str] = ALLOWED_ADAPTERS, allowed_backends: set[str] | frozenset[str] = ALLOWED_BACKENDS, allowed_approaches: set[str] | frozenset[str] = ALLOWED_APPROACHES) -> dict¶

Reject any adapter recommendation outside the canonical shape / vocabulary.

guide_floor(data_type: str = 'tabular', data_size: str = 'medium', goal: str = 'predict', expertise: str = 'beginner', notes: str = '') -> AutoMLGuideRecommendation¶

Deterministic decision-tree adapter selector (the honest floor).

MCP Tools¶