Config file

This page explains every field in the data_generation section of the YAML config.

The SDK reads a single YAML file that contains both data_generation and training. You typically run:

run_data_generation(config_path=...)

run_training(config_path=...)

File format

Your YAML file is a run config that includes top-level keys and a nested data_generation section:

run_steps: ["data_generation", "training"]
pipeline_provider: "local"
metric_name: "custom"

data_generation:
  # described below
  metric: {}
  source_data: {}
  llm: {}
  generation: {}
  output: {}
  labelling: {}
  data_quality_metrics: {}

Configuration structure

The data_generation section has six parts:

metric: what you’re generating data for (classes, rubrics, input format)
source_data: the seed dataset (CSV or Hugging Face) + sampling
llm: the primary LLM used for generation
generation: how many examples to generate, distribution, concurrency
output: where the generated dataset is written/published
labelling + data_quality_metrics: optional steps

`metric`

Defines the classification metric you are generating data for.

Field	Type	Required	Default	Notes
`metric.name`	`string`	Yes	—	Metric name used in prompts and artifacts.
`metric.description`	`string \| null`	Conditionally	`null`	Required when `llmaj_source_prompt` is empty.
`metric.type`	`string`	Yes	—	One of `binary`, `multi-class`.
`metric.input_format`	`string`	Yes	—	One of `single`, `tuple`, `rag`, `span_with_tools`, `trace`, or `session`. `trace` and `session` require label-only mode or skipping data generation.
`metric.class_labels`	`list[class_label] \| null`	Conditionally	`null`	Required when `llmaj_source_prompt` is empty.
`metric.llmaj_source_prompt`	`string \| null`	No	`null`	If set, description + class labels may be extracted automatically. We recommend you to set this

`metric.class_label`

Field	Type	Required	Notes
`name`	`string`	Yes	Human-readable label name (e.g. `positive`).
`label`	`int \| string \| bool`	Yes	The label value used in your dataset. Must be unique.
`rubric`	`string`	Conditionally	Required when `llmaj_source_prompt` is empty.

`source_data`

Defines where seed examples come from and how they’re sampled.

`source_data.dataset`

Field	Type	Required	Default	Notes
`source_type`	`string`	Yes	—	`huggingface` or `csv`.
`columns.features`	`list[string]`	Yes	—	Feature column names. Rules depend on `metric.input_format`.
`columns.label`	`string`	Yes*	`"label"` (base config)	Required unless `labelling.label_only_mode` is true.
`huggingface.name`	`string`	If HF	—	Dataset name like `org/repo`.
`huggingface.split`	`string`	If HF	—	Split name (e.g. `train`, `test`).
`huggingface.revision`	`string`	No	`""`	Optional Git revision/branch.
`csv.file_path`	`string`	If CSV	—	Path to a CSV file on disk.
`csv.train_file_path`	`string`	No	`""`	Used for label-only workflows.

`source_data.sampling`

Field	Type	Required	Default (base config)	Notes
`enhancement_fraction`	`float`	Yes	`0.2`	Must be between 0 and 1 (exclusive).
`max_enhancement_samples_per_class`	`int`	Yes	`50000`	Upper bound for seed examples per class.
`include_enhancement_in_train`	`bool`	Yes	`true`	If true, seed examples are included in the final train split.
`seed`	`int`	Yes	`42`	Controls sampling reproducibility.
`pinned_enhancement_row_ids`	`list[int]`	No	`[]`	Force-include specific rows in the seed set.

`llm`

Primary LLM used for generation.

Field	Type	Required	Default (base config)	Notes
`provider`	`string`	Yes	`"openai"`	One of `anthropic`, `azure`, `gemini`, `groq`, `openai`, or `vegas`.
`model`	`string`	Yes	`"gpt-4.1"`	Provider model identifier.
`embedding_model`	`string`	Yes	`"text-embedding-3-small"`	Used for optional quality metrics.

`generation`

Controls how many examples you generate and how fast.

Field	Type	Required	Default (base config)	Notes
`context_examples`	`int`	Yes	`5`	Number of seed examples placed in the prompt per call. For `rag` and `span_with_tools`, must be `1`.
`examples_per_call`	`int`	Yes	`10`	Examples requested per LLM call.
`total_examples`	`int`	Yes	`2000`	Total synthetic examples across all classes.
`label_distribution`	`map[string,float]`	Yes	`{}`	Empty means auto-infer from the holdout/test split; otherwise keys should sum to ~1.0.
`async_config.max_concurrent_requests`	`int`	Yes	`5`	Concurrency.
`debug_save_prompts`	`bool`	No	`false`	Save prompts/responses for debugging.
`special_instructions`	`string \| null`	No	`null`	Extra guidance appended to prompts.
`llm_provider`	`string \| null`	No	`null`	Optional override; must be set together with `llm_model`.
`llm_model`	`string \| null`	No	`null`	Optional override; must be set together with `llm_provider`.

Because the default is 5, a span_with_tools generation config must override it explicitly:

data_generation:
  metric:
    input_format: "span_with_tools"
  generation:
    context_examples: 1

`output`

Controls where the generated dataset goes.

Field	Type	Required	Default (base config)	Notes
`push_to_hub`	`bool`	Yes	`true`	Push to Hugging Face Hub.
`push_to_object_store`	`bool`	No	`false`	Upload a compressed dataset artifact to the configured object store.
`local_path`	`string`	Yes	`"./data/generated_data"`	Local artifact directory.
`object_store_bucket`	`string \| null`	If object store	`"${LUNA_OBJECT_STORE_BUCKET:-luna-ft}"`	Required when `push_to_object_store` is true.
`object_store_blob`	`string \| null`	If object store	`"generated_data"`	Required when `push_to_object_store` is true.
`dataset.repo_name`	`string`	Yes	—	Local folder name and/or hub repo name.
`dataset.private`	`bool`	Yes	`true`	Hub visibility.
`dataset.splits.train`	`string`	Yes	`"train"`	Output split name for training data.
`dataset.splits.test`	`string`	Yes	`"test"`	Output split name for holdout/test data.
`hub.organization`	`string`	Yes	`"rungalileo"`	Required if `push_to_hub` is true.

`labelling`

Optional: run LLM-as-a-judge labeling workflows. If enabled, metric.llmaj_source_prompt is required.

Field	Type	Required	Default	Notes
`enabled`	`bool`	No	`false`	Enables labeling.
`label_only_mode`	`bool`	No	`false`	Label an existing train file without generating new examples.

`data_quality_metrics`

Optional: compute UMAP and drift-based diagnostics.

Field	Type	Required	Default (base config)	Notes
`enabled`	`bool`	No	`true`	Master switch.
`random_seed`	`int`	No	`42`	Reproducibility.
`artifact_subdir`	`string`	No	`"artifacts"`	Folder under `output.local_path`.
`umap.enabled`	`bool`	No	`false`	Enable UMAP projections.
`umap.mode`	`string`	No	`"holdout_anchor"`	`joint` or `holdout_anchor`.
`umap.umap_neighbors`	`int`	No	`15`	Neighbor count for UMAP.
`umap.umap_min_dist`	`float`	No	`0.1`	Minimum distance for UMAP.
`drift.enabled`	`bool`	No	`true`	Enable drift scoring.
`drift.k_neighbors`	`int`	No	`30`	k-NN size.
`drift.drift_threshold`	`float`	No	`0.95`	Percentile threshold in [0,1].
`drift.index_dir`	`string`	No	`"/tmp/faiss"`	Directory for the FAISS index.
`drift.output_filename`	`string`	No	`"drift_scores.parquet"`	Output file for drift scores.
`drift.sort_by_score_desc`	`bool`	No	`true`	Sort drift results descending.

Overview

Get Started

Observability

Evaluation Metrics

Signals

AI Assistant

Luna Studio

Experiments

Agent Control

Annotations

Integrations

Security

References

File format

Configuration structure

`metric`

`metric.class_label`

`source_data`

`source_data.dataset`

`source_data.sampling`

`llm`

`generation`

`output`

`labelling`

`data_quality_metrics`

​File format

​Configuration structure

​metric

​metric.class_label

​source_data

​source_data.dataset

​source_data.sampling

​llm

​generation

​output

​labelling

​data_quality_metrics

File format

Configuration structure

`metric`

`metric.class_label`

`source_data`

`source_data.dataset`

`source_data.sampling`

`llm`

`generation`

`output`

`labelling`

`data_quality_metrics`