ICLEstimator.predict()

Submits an ICL inference job and returns a JobMonitor. Runs DFS (for RelationalDataset) or the flat pipeline (for TabularDataset), then ICL, on the task’s test split.

Parameters

Name	Type	Description	Optional
dataset	RelationalDataset | TabularDataset	Dataset to run inference on.	No
output_config	OutputConfig	Export destination. Use OutputConfig.snowflake() to write predictions to a Snowflake table. When None, predictions are not exported.	Yes
output_alias	str	Override the auto-generated output table name. When None, the name is derived from the dataset and task names.	Yes
overwrite_results	bool	If True, silently overwrite an existing Snowflake output table. If False (default), raise an error if the destination already exists.	Yes
test_batch_size	int	Number of rows per inference chunk. Larger values are faster but use more memory — the kv_cache used internally is memory-heavy, so this defaults to a conservative 256. Increase only if you know the target memory budget can absorb it.	Yes
materialize_results	bool	If True, block until the job completes and predictions are written to the export destination before returning. Requires output_config to be set. Defaults to True.	Yes
test_table	str	File/table path for the test split. Optional — falls back to the task’s own test_table when not provided.	Yes

Returns

A JobMonitor to poll status and retrieve the export path.

Requirements

• task.label_column must be set (see ICLEstimator).
• task.context_table must be set — ICL has no training loop, so context_table names the labeled data used directly as the zero-shot context.
• The task’s test_table (or the test_table argument) must exist and contain the required entity/time columns.
• If ICLConfig.sampling_strategy is "most_recent" or "mixed", the task must have a time_column set.

Examples

Relational dataset

from relationalai_predictive import ICLEstimator, ICLConfig, OutputConfig

model = ICLEstimator(connector=connector, config=ICLConfig(n_estimators=8))

output_config = OutputConfig.snowflake(
    database_name="DATABASE_NAME",
    schema_name="PUBLIC",
)

job = model.predict(
    relational_dataset,
    output_config=output_config,
    output_alias="EXPERIMENT_1",
)

By default, predictions are written to a table named PREDICTIONS_{output_alias} in the schema defined in the OutputConfig.

Flat / forecasting dataset

from relationalai_predictive import (
    ICLEstimator, ICLConfig, TabularTask, TabularDataset, TaskType, OutputConfig,
)

task = TabularTask(
    name="sales_forecast",
    task_type=TaskType.FORECASTING,
    label_column="sales",
    context_table="DATABASE.SCHEMA.SALES_CONTEXT",
    test_table="DATABASE.SCHEMA.SALES_TEST",
    time_column="date",
    prediction_length=12,
    max_context_length=64,
)
dataset = TabularDataset(connector=connector, name="sales", task=task)

model = ICLEstimator(connector=connector, config=ICLConfig(device="cpu"))

job = model.predict(
    dataset,
    output_config=OutputConfig.snowflake(database_name="DATABASE", schema_name="SCHEMA"),
    output_alias="SALES_FORECAST_1",
)

Non-blocking job

job = model.predict(
    relational_dataset,
    output_config=output_config,
    materialize_results=False,
)

# Later, once the job has finished:
job.materialize_results()

Common Validation Errors

• “ICLEstimator requires task.label_column to be set” — ICL has no training loop; the context table must carry labels.
• “ICL requires context_table to be set on the task” — context_table must be set explicitly.
• “ICLEstimator does not support task_type=… for RelationalDataset/TabularDataset” — see the supported task types table in ICLEstimator.
• “sampling_strategy=‘most_recent’/‘mixed’ requires a time_column on the task, but none is set” — either set time_column on the task or use sampling_strategy="stratified"/"random".
• “has no ‘test’ split” / “Inference requires ‘test_table’ to be set on the task” — set test_table= on the task, or pass test_table= to predict().
• “Table ’…’ is missing required column(s): […]” — the context table or test table is missing a required column (label, time, or entity column).