data-designer 0.9.2__tar.gz → 0.9.4__tar.gz

{data_designer-0.9.2 → data_designer-0.9.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: data-designer
-Version: 0.9.2
+Version: 0.9.4
 Summary: General framework for synthetic data generation
 License-Expression: Apache-2.0
 Classifier: Development Status :: 4 - Beta
@@ -15,14 +15,14 @@ Classifier: Programming Language :: Python :: 3.14
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Classifier: Topic :: Software Development
 Requires-Python: >=3.10
-Requires-Dist: data-designer-config==0.9.2
-Requires-Dist: data-designer-engine==0.9.2
+Requires-Dist: data-designer-config==0.9.4
+Requires-Dist: data-designer-engine==0.9.4
 Requires-Dist: huggingface-hub<2,>=1.0.1
 Requires-Dist: opentelemetry-api<1.44,>=1.43
 Requires-Dist: opentelemetry-exporter-prometheus<0.65,>=0.64b0
 Requires-Dist: opentelemetry-sdk<1.44,>=1.43
 Requires-Dist: packaging<27,>=25
-Requires-Dist: pandas<3,>=2.3.3
+Requires-Dist: pandas<4,>=2.3.3
 Requires-Dist: prometheus-client<1,>=0.24
 Requires-Dist: prompt-toolkit<4,>=3.0.0
 Requires-Dist: pyarrow<25,>=24
@@ -173,7 +173,7 @@ After installation, invoke the `data-designer` skill or describe the dataset you
 
 This repository supports agent-assisted development — see [CONTRIBUTING.md](CONTRIBUTING.md) for the recommended workflow.
 
-- **[Contributing Guide](CONTRIBUTING.md)** – How to contribute, including agent-assisted workflows
+- **[Contributing Guide](CONTRIBUTING.md)** – How to contribute, including issue triage before external PRs and agent-assisted workflows
 - **[GitHub Issues](https://github.com/NVIDIA-NeMo/DataDesigner/issues)** – Report bugs or make a feature request
 
 ---

{data_designer-0.9.2 → data_designer-0.9.4}/README.md

@@ -139,7 +139,7 @@ After installation, invoke the `data-designer` skill or describe the dataset you
 
 This repository supports agent-assisted development — see [CONTRIBUTING.md](CONTRIBUTING.md) for the recommended workflow.
 
-- **[Contributing Guide](CONTRIBUTING.md)** – How to contribute, including agent-assisted workflows
+- **[Contributing Guide](CONTRIBUTING.md)** – How to contribute, including issue triage before external PRs and agent-assisted workflows
 - **[GitHub Issues](https://github.com/NVIDIA-NeMo/DataDesigner/issues)** – Report bugs or make a feature request
 
 ---

{data_designer-0.9.2 → data_designer-0.9.4}/pyproject.toml

@@ -47,7 +47,7 @@ dependencies = [
     "opentelemetry-exporter-prometheus>=0.64b0,<0.65",
     "opentelemetry-sdk>=1.43,<1.44",
     "packaging>=25,<27",
-    "pandas>=2.3.3,<3",
+    "pandas>=2.3.3,<4",
     "prometheus-client>=0.24,<1",
     "prompt-toolkit>=3.0.0,<4",
     "pyarrow>=24,<25",

{data_designer-0.9.2 → data_designer-0.9.4}/src/data_designer/integrations/huggingface/client.py

@@ -119,7 +119,7 @@ class HuggingFaceHubClient:
             URL to the uploaded dataset
 
         Raises:
-            HuggingFaceUploadError: If validation fails or upload encounters errors
+            HuggingFaceHubClientUploadError: If validation fails or upload encounters errors
         """
         logger.info(f"🤗 Uploading dataset to Hugging Face Hub: {repo_id}")
 
@@ -162,7 +162,7 @@ class HuggingFaceHubClient:
             private: Whether to create private repo
 
         Raises:
-            HuggingFaceUploadError: If repository creation fails
+            HuggingFaceHubClientUploadError: If repository creation fails
         """
         logger.info(f"{LOG_INDENT}{RandomEmoji.working()} Checking if repository exists...")
         try:
@@ -203,7 +203,7 @@ class HuggingFaceHubClient:
             parquet_folder: Path to folder containing parquet files
 
         Raises:
-            HuggingFaceUploadError: If upload fails
+            HuggingFaceHubClientUploadError: If upload fails
         """
         logger.info(f"{LOG_INDENT}{RandomEmoji.loading()} Uploading main dataset files...")
         try:
@@ -225,7 +225,7 @@ class HuggingFaceHubClient:
             images_folder: Path to images folder
 
         Raises:
-            HuggingFaceUploadError: If upload fails
+            HuggingFaceHubClientUploadError: If upload fails
         """
         if not images_folder.exists():
             return
@@ -255,7 +255,7 @@ class HuggingFaceHubClient:
             processors_folder: Path to folder containing processor output directories
 
         Raises:
-            HuggingFaceUploadError: If upload fails
+            HuggingFaceHubClientUploadError: If upload fails
         """
         if not processors_folder.exists():
             return
@@ -290,7 +290,7 @@ class HuggingFaceHubClient:
             builder_config_path: Path to builder_config.json file
 
         Raises:
-            HuggingFaceUploadError: If upload fails
+            HuggingFaceHubClientUploadError: If upload fails
         """
         logger.info(f"{LOG_INDENT}{RandomEmoji.loading()} Uploading configuration files...")
 
@@ -345,7 +345,7 @@ class HuggingFaceHubClient:
             tags: Additional custom tags for the dataset
 
         Raises:
-            HuggingFaceUploadError: If dataset card generation or upload fails
+            HuggingFaceHubClientUploadError: If dataset card generation or upload fails
         """
         try:
             with open(metadata_path) as f:
@@ -455,7 +455,7 @@ class HuggingFaceHubClient:
             base_dataset_path: Path to dataset directory
 
         Raises:
-            HuggingFaceUploadError: If directory structure is invalid
+            HuggingFaceHubClientUploadError: If directory structure is invalid
         """
         if not base_dataset_path.exists():
             raise HuggingFaceHubClientUploadError(f"Dataset path does not exist: {base_dataset_path}")

{data_designer-0.9.2 → data_designer-0.9.4}/src/data_designer/interface/__init__.py

@@ -22,19 +22,41 @@ if TYPE_CHECKING:
         DataDesignerWorkflowError,
     )
     from data_designer.interface.results import DatasetCreationResults  # noqa: F401
+    from data_designer.interface.workflow_metadata import (  # noqa: F401
+        CompletedWorkflowStageMetadata,
+        FailedWorkflowStageMetadata,
+        RunningWorkflowStageMetadata,
+        SkippedWorkflowStageMetadata,
+        WorkflowMetadata,
+        WorkflowStageMetadata,
+        validate_workflow_stage_metadata,
+    )
 
 _LAZY_IMPORTS: dict[str, tuple[str, str]] = {
     "CompositeWorkflow": ("data_designer.interface.composite_workflow", "CompositeWorkflow"),
     "CompositeWorkflowResults": ("data_designer.interface.composite_workflow", "CompositeWorkflowResults"),
+    "CompletedWorkflowStageMetadata": (
+        "data_designer.interface.workflow_metadata",
+        "CompletedWorkflowStageMetadata",
+    ),
     "DataDesigner": ("data_designer.interface.data_designer", "DataDesigner"),
     "DataDesignerEarlyShutdownError": ("data_designer.interface.errors", "DataDesignerEarlyShutdownError"),
     "DataDesignerGenerationError": ("data_designer.interface.errors", "DataDesignerGenerationError"),
     "DataDesignerProfilingError": ("data_designer.interface.errors", "DataDesignerProfilingError"),
     "DataDesignerWorkflowError": ("data_designer.interface.errors", "DataDesignerWorkflowError"),
     "DatasetCreationResults": ("data_designer.interface.results", "DatasetCreationResults"),
+    "FailedWorkflowStageMetadata": ("data_designer.interface.workflow_metadata", "FailedWorkflowStageMetadata"),
     "ResumeMode": ("data_designer.config.run_config", "ResumeMode"),
+    "RunningWorkflowStageMetadata": ("data_designer.interface.workflow_metadata", "RunningWorkflowStageMetadata"),
     "SkippedStageResult": ("data_designer.interface.composite_workflow", "SkippedStageResult"),
     "SkippedStageStatus": ("data_designer.interface.composite_workflow", "SkippedStageStatus"),
+    "SkippedWorkflowStageMetadata": ("data_designer.interface.workflow_metadata", "SkippedWorkflowStageMetadata"),
+    "WorkflowMetadata": ("data_designer.interface.workflow_metadata", "WorkflowMetadata"),
+    "WorkflowStageMetadata": ("data_designer.interface.workflow_metadata", "WorkflowStageMetadata"),
+    "validate_workflow_stage_metadata": (
+        "data_designer.interface.workflow_metadata",
+        "validate_workflow_stage_metadata",
+    ),
 }
 
 __all__ = list(_LAZY_IMPORTS.keys())

{data_designer-0.9.2 → data_designer-0.9.4}/src/data_designer/interface/composite_workflow.py

@@ -40,6 +40,7 @@ from data_designer.interface.results import (
     _export_jsonl,
     _export_parquet,
 )
+from data_designer.interface.workflow_metadata import WorkflowMetadata, validate_workflow_stage_metadata
 
 if TYPE_CHECKING:
     import pandas as pd
@@ -219,8 +220,19 @@ class CompositeWorkflow:
         _validate_dir_name(name, "stage name")
         if any(stage.name == name for stage in self._stages):
             raise DataDesignerWorkflowError(f"Stage name {name!r} is already used in workflow {self.name!r}.")
-        if num_records is not None and num_records < 1:
-            raise DataDesignerWorkflowError("Stage num_records must be at least 1.")
+        if num_records is not None:
+            if not isinstance(num_records, int) or isinstance(num_records, bool):
+                raise DataDesignerWorkflowError("Stage num_records must be an integer.")
+            if num_records < 1:
+                raise DataDesignerWorkflowError("Stage num_records must be at least 1.")
+        if on_success_version is not None and not isinstance(on_success_version, str):
+            raise DataDesignerWorkflowError("Stage on_success_version must be a string.")
+        if not isinstance(allow_empty, bool):
+            raise DataDesignerWorkflowError("Stage allow_empty must be a boolean.")
+        if not isinstance(sampling_strategy, SamplingStrategy):
+            raise DataDesignerWorkflowError("Stage sampling_strategy must be a SamplingStrategy.")
+        if selection_strategy is not None and not isinstance(selection_strategy, (IndexRange, PartitionBlock)):
+            raise DataDesignerWorkflowError("Stage selection_strategy must be an IndexRange or PartitionBlock.")
         _validate_stage_output(output)
         output_processors = output_processors or []
         _validate_distinct_output_processors(config_builder, output_processors)
@@ -276,6 +288,7 @@ class CompositeWorkflow:
         workflow_path.mkdir(parents=True, exist_ok=True)
         prior_metadata = _read_prior_workflow_metadata(workflow_path, self.name, resume)
         metadata: dict[str, Any] = {
+            **(prior_metadata or {}),
             "name": self.name,
             "library_version": get_library_version(),
             "stages": [],
@@ -388,6 +401,10 @@ class CompositeWorkflow:
                     f"Cannot resume workflow {self.name!r}: stage {stage.name!r} is not reusable."
                 )
 
+            if stage_resume == ResumeMode.ALWAYS and prior_stage_metadata is not None:
+                prior_stage = validate_workflow_stage_metadata(prior_stage_metadata)
+                stage_metadata.update(prior_stage.model_extra or {})
+
             if stage_resume == ResumeMode.NEVER and stage_path.exists():
                 shutil.rmtree(stage_path)
 
@@ -598,6 +615,22 @@ def _read_prior_workflow_metadata(
         raise DataDesignerWorkflowError(
             f"Cannot resume workflow {workflow_name!r}: workflow metadata has invalid shape."
         )
+    try:
+        validated_metadata = WorkflowMetadata.model_validate(metadata)
+    except ValidationError as exc:
+        if resume != ResumeMode.ALWAYS:
+            error = exc.errors(include_url=False, include_input=False)[0]
+            location = ".".join(str(part) for part in error["loc"])
+            logger.warning(
+                "Workflow metadata for %r has invalid field %s (%s); starting fresh.",
+                workflow_name,
+                location,
+                error["msg"],
+            )
+            return None
+        raise DataDesignerWorkflowError(
+            f"Cannot resume workflow {workflow_name!r}: workflow metadata has invalid shape."
+        ) from exc
     if metadata.get("name") != workflow_name:
         if resume != ResumeMode.ALWAYS:
             logger.warning("Workflow metadata for %r has a different name; starting fresh.", workflow_name)
@@ -605,7 +638,7 @@ def _read_prior_workflow_metadata(
         raise DataDesignerWorkflowError(
             f"Cannot resume workflow {workflow_name!r}: workflow metadata name does not match."
         )
-    return metadata
+    return validated_metadata.model_dump(mode="json", exclude_unset=True)
 
 
 def _get_prior_stage_metadata(
@@ -870,9 +903,13 @@ def _parquet_files(path: Path) -> list[Path]:
 def _write_workflow_metadata(workflow_path: Path, metadata: dict[str, Any]) -> None:
     path = workflow_path / WORKFLOW_METADATA_FILENAME
     tmp_path = path.with_name(f"{path.name}.tmp.{os.getpid()}.{uuid.uuid4().hex}")
+    try:
+        validated_metadata = WorkflowMetadata.model_validate(metadata)
+    except ValidationError as exc:
+        raise DataDesignerWorkflowError("Workflow metadata has invalid shape.") from exc
     try:
         with tmp_path.open("w", encoding="utf-8") as f:
-            json.dump(metadata, f, indent=2, sort_keys=True)
+            json.dump(validated_metadata.model_dump(mode="json", exclude_unset=True), f, indent=2, sort_keys=True)
             f.flush()
             os.fsync(f.fileno())
         os.replace(tmp_path, path)

{data_designer-0.9.2 → data_designer-0.9.4}/src/data_designer/interface/data_designer.py

@@ -354,11 +354,10 @@ class DataDesigner(DataDesignerInterface[DatasetCreationResults]):
             )
 
         try:
-            profiler_column_configs = config_builder.get_column_configs() or builder.data_designer_config.columns
             profiler = self._create_dataset_profiler(
                 config_builder,
                 resource_provider,
-                column_configs=profiler_column_configs,
+                column_configs=builder.data_designer_config.columns,
             )
             analysis = profiler.profile_dataset(num_records, dataset_for_profiler)
         except Exception as e:
@@ -457,7 +456,11 @@ class DataDesigner(DataDesignerInterface[DatasetCreationResults]):
             dataset_for_profiler = processed_dataset
 
         try:
-            profiler = self._create_dataset_profiler(config_builder, resource_provider)
+            profiler = self._create_dataset_profiler(
+                config_builder,
+                resource_provider,
+                column_configs=builder.data_designer_config.columns,
+            )
             analysis = profiler.profile_dataset(num_records, dataset_for_profiler)
         except Exception as e:
             raise DataDesignerProfilingError(f"🛑 Error profiling preview dataset: {e}") from e
@@ -623,11 +626,7 @@ class DataDesigner(DataDesignerInterface[DatasetCreationResults]):
         """Get the runtime configuration applied to dataset generation.
 
         Returns:
-            The active RunConfig instance. Note that ``RunConfig`` normalizes
-            some fields on construction (e.g., ``shutdown_error_rate`` becomes
-            ``1.0`` when ``disable_early_shutdown=True``), so the returned
-            object may not exactly equal the one originally passed to
-            ``set_run_config``.
+            The active RunConfig instance.
         """
         return self._run_config
 

data_designer-0.9.4/src/data_designer/interface/workflow_metadata.py

@@ -0,0 +1,97 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from typing import Annotated, Any, Literal, TypeAlias
+
+from pydantic import BaseModel, ConfigDict, Field, TypeAdapter
+
+
+class _WorkflowStageMetadataBase(BaseModel):
+    """Common metadata for a composite workflow stage."""
+
+    model_config = ConfigDict(extra="allow")
+
+    status: Literal["running", "failed", "completed", "completed_empty", "skipped_empty_upstream"]
+    index: int
+    name: str
+    stage_dir: str
+    depends_on: list[str]
+    allow_empty: bool
+    on_success_version: str | None
+    output_processors: list[dict[str, Any]]
+    output: str
+    sampling_strategy: str
+    selection_strategy: dict[str, Any] | None
+
+
+class _StartedWorkflowStageMetadata(_WorkflowStageMetadataBase):
+    fingerprint: str
+    num_records_requested: int
+    seeded_from_stage: str | None
+    seed_path: str | None
+    config: dict[str, Any]
+
+
+class RunningWorkflowStageMetadata(_StartedWorkflowStageMetadata):
+    """Metadata for a running workflow stage."""
+
+    status: Literal["running"]
+
+
+class FailedWorkflowStageMetadata(_StartedWorkflowStageMetadata):
+    """Metadata for a failed workflow stage."""
+
+    status: Literal["failed"]
+    duration_sec: float | None = None
+
+
+class CompletedWorkflowStageMetadata(_StartedWorkflowStageMetadata):
+    """Metadata for a completed workflow stage."""
+
+    status: Literal["completed", "completed_empty"]
+    num_records_actual: int
+    output_records: int
+    output_seed_path: str
+    callback_output_path: str | None
+    stage_output_override_path: str | None = None
+    output_processor_output_path: str | None
+    duration_sec: float
+
+
+class SkippedWorkflowStageMetadata(_WorkflowStageMetadataBase):
+    """Metadata for a stage skipped after an empty upstream stage."""
+
+    status: Literal["skipped_empty_upstream"]
+    upstream_stage: str
+
+
+WorkflowStageMetadata: TypeAlias = Annotated[
+    RunningWorkflowStageMetadata
+    | FailedWorkflowStageMetadata
+    | CompletedWorkflowStageMetadata
+    | SkippedWorkflowStageMetadata,
+    Field(discriminator="status"),
+]
+
+_WORKFLOW_STAGE_METADATA_ADAPTER = TypeAdapter(WorkflowStageMetadata)
+
+
+def validate_workflow_stage_metadata(value: object) -> WorkflowStageMetadata:
+    """Validate the stage union and return its concrete status model.
+
+    Use the concrete status models for runtime type checks.
+    """
+
+    return _WORKFLOW_STAGE_METADATA_ADAPTER.validate_python(value)
+
+
+class WorkflowMetadata(BaseModel):
+    """Metadata persisted for a composite workflow run."""
+
+    model_config = ConfigDict(extra="allow")
+
+    name: str
+    library_version: str
+    stages: list[WorkflowStageMetadata]

{data_designer-0.9.2 → data_designer-0.9.4}/tests/cli/commands/test_create_command.py

@@ -3,11 +3,20 @@
 
 from __future__ import annotations
 
+from pathlib import Path
 from unittest.mock import MagicMock, patch
 
+import pytest
+from typer.testing import CliRunner
+
 from data_designer.cli.commands.create import create_command
+from data_designer.cli.main import app
+from data_designer.config.run_config import RunConfig
 from data_designer.engine.storage.artifact_storage import ResumeMode
 
+runner = CliRunner()
+_CTRL = "data_designer.cli.controllers.generation_controller"
+
 # ---------------------------------------------------------------------------
 # create_command delegation tests
 # ---------------------------------------------------------------------------
@@ -229,3 +238,48 @@ def test_create_command_passes_tui_override(mock_ctrl_cls: MagicMock) -> None:
         tui=False,
         script_args=None,
     )
+
+
+# ---------------------------------------------------------------------------
+# create --run-config overlay tests
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("yaml_text", "expected_rate", "expected_effective_rate"),
+    [
+        ("disable_early_shutdown: true\n", 0.3, 1.0),
+        ("disable_early_shutdown: true\nshutdown_error_rate: 0.45\n", 0.45, 1.0),
+        ("shutdown_error_rate: 0.45\n", 0.45, 0.45),
+    ],
+    ids=["flag-only-keeps-baseline-rate", "flag-with-explicit-rate", "rate-only"],
+)
+@patch(f"{_CTRL}.DataDesigner")
+@patch(f"{_CTRL}.load_config_builder")
+def test_create_run_config_overlay_preserves_shutdown_error_rate(
+    mock_load_config: MagicMock,
+    mock_dd_cls: MagicMock,
+    tmp_path: Path,
+    yaml_text: str,
+    expected_rate: float,
+    expected_effective_rate: float,
+) -> None:
+    """`create --run-config` overlays the YAML file without discarding the baseline error rate."""
+    run_config_file = tmp_path / "run-config.yaml"
+    run_config_file.write_text(yaml_text)
+
+    mock_load_config.return_value = MagicMock()
+    mock_results = MagicMock()
+    mock_results.count_records.return_value = 10
+    mock_results.load_analysis.return_value = None
+    mock_dd = MagicMock()
+    mock_dd.run_config = RunConfig(shutdown_error_rate=0.3)
+    mock_dd.create.return_value = mock_results
+    mock_dd_cls.return_value = mock_dd
+
+    result = runner.invoke(app, ["create", "dataset.yaml", "--run-config", str(run_config_file)])
+
+    assert result.exit_code == 0, result.output
+    effective = mock_dd.set_run_config.call_args.args[0]
+    assert effective.shutdown_error_rate == expected_rate
+    assert effective.effective_shutdown_error_rate == expected_effective_rate

{data_designer-0.9.2 → data_designer-0.9.4}/tests/cli/controllers/test_generation_controller.py

@@ -866,7 +866,7 @@ def test_run_create_applies_run_config_precedence(
         }
     )
     mock_dd = MagicMock()
-    mock_dd.run_config = RunConfig(buffer_size=64, progress_interval=11.0, display_tui=True)
+    mock_dd.run_config = RunConfig(buffer_size=64, progress_interval=11.0, display_tui=True, shutdown_error_rate=0.3)
     mock_dd.create.return_value = _make_mock_create_results()
     mock_dd_cls.return_value = mock_dd
 
@@ -886,7 +886,8 @@ def test_run_create_applies_run_config_precedence(
     assert effective.buffer_size == 1000
     assert effective.progress_interval == 11.0
     assert effective.preserve_dropped_columns is False
-    assert effective.shutdown_error_rate == 1.0
+    assert effective.shutdown_error_rate == 0.3
+    assert effective.effective_shutdown_error_rate == 1.0
     assert effective.display_tui is False
     assert [item[0] for item in mock_dd.method_calls] == ["set_run_config", "create"]
     assert "Run config: run.yaml" in capsys.readouterr().out

{data_designer-0.9.2 → data_designer-0.9.4}/tests/cli/utils/test_config_loader.py

@@ -68,13 +68,23 @@ def test_load_run_config_accepts_empty_mapping(tmp_path: Path) -> None:
     assert loaded.model_dump(exclude_unset=True) == {}
 
 
-def test_load_run_config_preserves_explicit_partial_fields(tmp_path: Path) -> None:
+@pytest.mark.parametrize(
+    ("yaml_text", "expected"),
+    [
+        ("buffer_size: 250\ndisplay_tui: true\n", {"buffer_size": 250, "display_tui": True}),
+        ("disable_early_shutdown: true\n", {"disable_early_shutdown": True}),
+    ],
+    ids=["buffer-size-and-tui", "disable-early-shutdown-only"],
+)
+def test_load_run_config_preserves_explicit_partial_fields(
+    tmp_path: Path, yaml_text: str, expected: dict[str, object]
+) -> None:
     run_config_file = tmp_path / "run-config.yaml"
-    run_config_file.write_text("buffer_size: 250\ndisplay_tui: true\n")
+    run_config_file.write_text(yaml_text)
 
     loaded = load_run_config(str(run_config_file))
 
-    assert loaded.model_dump(exclude_unset=True) == {"buffer_size": 250, "display_tui": True}
+    assert loaded.model_dump(exclude_unset=True) == expected
 
 
 def test_load_run_config_preserves_partial_nested_fields(tmp_path: Path) -> None:

data_designer-0.9.4/tests/docs/test_chart_qa_recipe.py

@@ -0,0 +1,136 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import importlib.util
+from pathlib import Path
+from types import ModuleType
+from unittest.mock import MagicMock
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[4]
+RECIPE_PATH = REPO_ROOT / "docs" / "assets" / "recipes" / "chart_qa" / "chart_qa.py"
+
+
+def _load_recipe() -> ModuleType:
+    spec = importlib.util.spec_from_file_location("chart_qa_recipe", RECIPE_PATH)
+    assert spec is not None
+    assert spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+@pytest.fixture(scope="module")
+def recipe() -> ModuleType:
+    return _load_recipe()
+
+
+@pytest.fixture
+def seed_path(tmp_path: Path) -> Path:
+    path = tmp_path / "chart_seed.jsonl"
+    path.write_text('{"image_url": "https://example.com/chart.png"}\n', encoding="utf-8")
+    return path
+
+
+@pytest.mark.parametrize("prompt_version", ["level1", "level2"])
+def test_chart_prompt_versions_match_pipeline_config(recipe: ModuleType, seed_path: Path, prompt_version: str) -> None:
+    config = recipe.build_config(seed_path=str(seed_path), prompt_version=prompt_version).build()
+    columns = {column.name: column for column in config.columns}
+    models = {model.alias: model for model in config.model_configs}
+
+    assert list(columns) == ["question_type", "question", "answer", "quality_check"]
+    assert columns["question_type"].params.values == list(recipe.QUESTION_TYPES)
+    assert columns["question_type"].params.weights == list(recipe.QUESTION_TYPE_WEIGHTS)
+    assert columns["question"].model_alias == "chart-vlm-question"
+    assert columns["answer"].model_alias == "chart-vlm"
+    assert columns["quality_check"].model_alias == "chart-vlm-quality"
+    assert columns["answer"].extract_reasoning_content is True
+    assert columns["question"].multi_modal_context[0].column_name == "image_url"
+    assert columns["question"].multi_modal_context[0].data_type == "url"
+    assert columns["question"].multi_modal_context[0].image_format is None
+    assert columns["answer"].multi_modal_context == columns["question"].multi_modal_context
+    assert columns["quality_check"].multi_modal_context == columns["question"].multi_modal_context
+    assert "{{ answer__reasoning_content }}" in columns["quality_check"].prompt
+    assert config.seed_config.sampling_strategy == recipe.dd.SamplingStrategy.ORDERED
+    assert list(models) == ["chart-vlm", "chart-vlm-question", "chart-vlm-quality"]
+
+    for model in models.values():
+        inference = model.inference_parameters
+        assert model.model == recipe.DEFAULT_VLM_MODEL
+        assert model.provider == recipe.VLLM_PROVIDER_NAME
+        assert model.skip_health_check is False
+        assert inference.max_tokens == 50000
+        assert inference.timeout == 1200
+        assert inference.temperature == 0.6
+        assert inference.top_p == 0.95
+        assert inference.max_parallel_requests == 20
+        assert inference.extra_body == recipe.PRODUCTION_SAMPLING
+
+
+@pytest.mark.parametrize(
+    ("prompt_version", "raw_prompts"),
+    [
+        (
+            "level1",
+            ("LEVEL1_QUESTION_PROMPT", "LEVEL1_ANSWER_PROMPT", "LEVEL1_QUALITY_PROMPT"),
+        ),
+        (
+            "level2",
+            (
+                "LEVEL2_QUESTION_PROMPT",
+                "LEVEL2_ANSWER_PROMPT",
+                "LEVEL2_QUALITY_PROMPT",
+            ),
+        ),
+    ],
+)
+def test_prompt_pack_matches_runtime_whitespace_handling(
+    recipe: ModuleType, prompt_version: str, raw_prompts: tuple[str, str, str]
+) -> None:
+    prompt_pack = recipe.get_prompt_pack(prompt_version)
+
+    assert tuple(prompt_pack) == tuple(getattr(recipe, name).rstrip() for name in raw_prompts)
+    assert all(not prompt.endswith(("\n", "\r", " ", "\t")) for prompt in prompt_pack)
+
+    if prompt_version == "level2":
+        assert "DIFFICULTY FLOOR (required)" in prompt_pack.question
+        assert "NO-ATTEMPT HANDLING" in prompt_pack.answer
+
+
+def test_unknown_prompt_version_is_rejected(recipe: ModuleType) -> None:
+    with pytest.raises(ValueError, match="Unknown prompt version"):
+        recipe.get_prompt_pack("level3")
+
+
+def test_create_dataset_uses_production_provider_settings(
+    recipe: ModuleType, seed_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    config_builder = recipe.build_config(seed_path=str(seed_path))
+    dataset_results = MagicMock()
+    data_designer = MagicMock()
+    data_designer.create.return_value = dataset_results
+    data_designer_class = MagicMock(return_value=data_designer)
+    monkeypatch.setattr(recipe, "DataDesigner", data_designer_class)
+
+    result = recipe.create_dataset(
+        config_builder,
+        num_records=7,
+        vllm_endpoint="http://localhost:8000/v1",
+        prompt_version="level1",
+        artifact_path="artifacts",
+    )
+
+    assert result is dataset_results
+    provider = data_designer_class.call_args.kwargs["model_providers"][0]
+    assert provider.name == "vllm-vl"
+    assert provider.endpoint == "http://localhost:8000/v1"
+    assert provider.provider_type == "openai"
+    assert provider.api_key is None
+    data_designer.create.assert_called_once_with(
+        config_builder,
+        num_records=7,
+        dataset_name="chart_qa_level1",
+    )