Skip to content

Adapt Autosklearn¶

Optional MLJAR AutoML adapter for the MVP workspace. MCP exposes 11 tools.

Cluster: Uncategorised | Type: component | MCP Tools: 15

Overview¶

MLJAR-first AutoML adapter kept under the historical adapt_autosklearn name for compatibility. The preferred backend is MLJAR (mljar-supervised); auto-sklearn naming and execution are compatibility-only. When MLJAR is unavailable, the component surfaces that dependency gap through capability discovery, structured degradation, and warning cards. Non-production requests may still use the deterministic majority-classifier (classification) or mean-regressor (regression) baseline; production mode blocks that fallback.

When to use:

When you want MLJAR AutoML search, structured reports, artifact manifests, and explicit dependency health through G6's AIBlock interface
On any platform when you want a deterministic baseline (majority class / mean) with a loud qualified-draft reliability envelope
As a sanity-check baseline in adapt_automl benchmarking -- the fallback's accuracy bounds the real AutoML's value-add

Example:

from mvp.adapt_autosklearn import AdaptAutoSklearnBlock, AutoSklearnInput

block = AdaptAutoSklearnBlock(name="autosklearn")
result = block.infer(AutoSklearnInput(
    X=[[1.0, 2.0], [2.0, 3.0], [3.0, 4.0]],  # n_samples x n_features
    y=[0, 0, 1],                              # class labels for classification
    task="classification",                    # or "regression" / "auto"
    mode="fit_predict",                       # or "fit" / "predict"
    time_left_for_this_task=60,               # MLJAR total search budget (sec)
    allow_fallback=True,                      # production_mode=True still blocks fallback
))
# With MLJAR installed:
#   result.value.backend == "mljar"; reports and artifact_manifest are populated
# With MLJAR absent in non-production mode:
#   result.value.backend == "fallback"
#   result.value.best_model == "majority_classifier"
#   result.value.predictions == [0, 0, 0]
#   result.value.score == 0.6667  (2/3 labels were 0)
#   result.value.completion_state == "qualified-draft"
#   result.value.warning_card explains that real AutoML search did not run

Works well with: adapt_automl (richer FLAML/AutoGluon/H2O AutoML when available), adapt_algo_selector (rule-based algorithm picker as a faster alternative), adapt_sklearn (manual model selection)

Public API¶

`AdaptAutoSklearnBlock(AIBlock[AutoSklearnInput, AutoSklearnOutput, Any])`¶

Stateful optional adapter for auto-sklearn classifiers and regressors.

Field	Type	Default
`name`	`str`	'adapt_autosklearn'
`state`	`Any \\| None`	None

Methods:

`infer(data: AutoSklearnInput) -> Result[AutoSklearnOutput]`¶

`bias() -> dict[str, Any]`¶

`AutoSklearnInput(BaseModel)`¶

Input for fitting or using the AutoML compatibility adapter.

Field	Type	Default
`X`	`list[list[float]]`	required
`y`	`list[float \\| int \\| str]`	Field(default_factory=list)
`backend`	`Literal['auto', 'mljar', 'autosklearn']`	'auto'
`mode`	`Literal['fit', 'predict', 'fit_predict']`	'fit_predict'
`task`	`Literal['auto', 'classification', 'regression']`	'auto'
`mljar_mode`	`Literal['Explain', 'Perform', 'Compete', 'Optuna']`	'Explain'
`validation_strategy`	`str \\| dict[str, Any]`	'auto'
`explain_level`	`int`	Field(default=2, ge=0, le=2)
`results_path`	`str`	''
`eval_metric`	`str`	'auto'
`time_left_for_this_task`	`int`	Field(default=30, ge=1, le=86400)
`per_run_time_limit`	`int`	Field(default=10, ge=1, le=86400)
`ensemble_size`	`int`	Field(default=1, ge=0, le=500)
`seed`	`int`	Field(default=42, ge=0)
`n_jobs`	`int`	Field(default=1, ge=1, le=64)
`include_estimators`	`list[str]`	Field(default_factory=list)
`exclude_estimators`	`list[str]`	Field(default_factory=list)
`allow_fallback`	`bool`	True
`production_mode`	`bool`	False
`request_id`	`str`	''
`task_id`	`str`	''
`run_id`	`str`	''

`AutoSklearnLeaderboardEntry(BaseModel)`¶

One candidate entry reported by an AutoML backend.

Field	Type	Default
`rank`	`int`	required
`name`	`str`	required
`score`	`float \\| None`	None
`params`	`dict[str, Any]`	Field(default_factory=dict)

`DegradationNotice(BaseModel)`¶

Normalized degradation notice for missing or partial backend capability.

Field	Type	Default
`reason`	`str`	required
`missing`	`list[str]`	Field(default_factory=list)
`impact`	`str`	''
`install_hint`	`str`	''
`recoverable`	`bool`	True

`BackendNativeOpRequest(BaseModel)`¶

Constrained request for an allowlisted backend-native operation.

Field	Type	Default
`operation`	`str`	required
`input`	`dict[str, Any]`	Field(default_factory=dict)
`safety_context`	`dict[str, Any]`	Field(default_factory=dict)
`dry_run`	`bool`	False
`max_runtime_ms`	`int`	Field(default=10000, ge=1, le=300000)

`CapabilityDiscovery(BaseModel)`¶

AutoML backend capability discovery payload.

Field	Type	Default
`component`	`str`	'adapt_autosklearn'
`backend`	`str`	'mljar'
`available`	`bool`	False
`status`	`str`	'missing_dependency'
`supported_modes`	`list[str]`	Field(default_factory=list)
`supported_ops`	`list[str]`	Field(default_factory=list)
`degraded_ops`	`list[str]`	Field(default_factory=list)
`safety_policy`	`dict[str, Any]`	Field(default_factory=dict)

`AutoSklearnOutput(BaseModel)`¶

Result returned by AdaptAutoSklearnBlock.

Field	Type	Default
`task`	`str`	required
`backend`	`str`	required
`best_model`	`str`	required
`predictions`	`list[Any] \\| None`	None
`score`	`float \\| None`	None
`leaderboard`	`list[AutoSklearnLeaderboardEntry]`	Field(default_factory=list)
`degraded`	`bool`	False
`degradation_reason`	`str`	''
`degradation`	`list[dict[str, Any]]`	Field(default_factory=list)
`warning_card`	`dict[str, Any]`	Field(default_factory=dict)
`available_with`	`list[str]`	Field(default_factory=list)
`fallback_used`	`bool`	False
`completion_state`	`Literal['verified', 'qualified-draft', 'blocked-escalated']`	'qualified-draft'
`evidence`	`dict[str, Any]`	Field(default_factory=dict)
`request_id`	`str`	''
`task_id`	`str`	''
`run_id`	`str`	''
`report`	`dict[str, Any]`	Field(default_factory=dict)
`artifact_manifest`	`dict[str, Any]`	Field(default_factory=dict)
`explainability`	`dict[str, Any]`	Field(default_factory=dict)
`model_cards`	`list[dict[str, Any]]`	Field(default_factory=list)
`warnings`	`list[str]`	Field(default_factory=list)
`agentic_evidence`	`dict[str, Any]`	Field(default_factory=dict)

MCP Tools¶

Operation	Source
`ops`	adapt_autosklearn_mcp
`help`	adapt_autosklearn_mcp
`readiness_check`	adapt_autosklearn_mcp
`list_strategies`	adapt_autosklearn_mcp
`infer_task`	adapt_autosklearn_mcp
`decide_missing_dependency`	adapt_autosklearn_mcp
`explain_fallback_model`	adapt_autosklearn_mcp
`list_patterns`	adapt_autosklearn_mcp
`discover_capabilities`	adapt_autosklearn_mcp
`dependency_health`	adapt_autosklearn_mcp
`structured_report`	adapt_autosklearn_mcp
`artifact_manifest`	adapt_autosklearn_mcp
`backend_native_op`	adapt_autosklearn_mcp
`classification`	adapt_autosklearn_mcp
`regression`	adapt_autosklearn_mcp