generate

generate

Classes:

Name	Description
ValidationParameters	Configuration for record and sequence validation.
StructuredGenerationParameters	Configuration for vLLM structured generation.
GenerateParameters	Configuration parameters for synthetic data generation.

Functions:

Name	Description
resolve_structured_generation_schema_method	Resolve auto schema method from the configured structured-output backend.
structural_tag_backend_error_message	Return an error message when backend cannot serve structural_tag.

ValidationParameters pydantic-model

Bases: Parameters, BaseModel

Configuration for record and sequence validation.

These parameters control the validation and automatic fixes when going from LLM output to tabular data.

Fields:

• group_by_accept_no_delineator (bool)
• group_by_ignore_invalid_records (bool)
• group_by_fix_non_unique_value (bool)
• group_by_fix_unordered_records (bool)

group_by_accept_no_delineator pydantic-field

Whether to accept completions without both beginning and end of sequence delineators as a single sequence.

group_by_ignore_invalid_records pydantic-field

Whether to ignore invalid records in a sequence and proceed with the valid records.

group_by_fix_non_unique_value pydantic-field

Whether to automatically fix non-unique group-by values in a sequence by using the first unique value for all records.

group_by_fix_unordered_records pydantic-field

Whether to automatically fix unordered records in a sequence by sorting the records.

StructuredGenerationParameters pydantic-model

Bases: Parameters, BaseModel

Configuration for vLLM structured generation.

These parameters control whether generation is constrained to schema-shaped output, which backend enforces the constraint, and how the constraint schema is built.

Fields:

• enabled (bool)
• backend (StructuredGenerationBackend)
• schema_method (StructuredGenerationSchemaMethod)
• use_single_sequence (bool)

Validators:

• _validate_structural_tag_backend

enabled pydantic-field

Whether to use structured generation for better format control.

backend pydantic-field

The backend used by vLLM when structured generation is enabled. Supported backends: 'outlines', 'guidance', 'xgrammar', 'lm-format-enforcer'. 'auto' will allow vLLM to choose the backend.

schema_method pydantic-field

The method used to generate the schema from your dataset and pass it to the generation backend. 'auto' picks 'structural_tag' on xgrammar-capable backends and 'regex' otherwise. 'regex' uses a custom regex construction method that tends to be more comprehensive than 'json_schema' at the cost of speed. 'structural_tag' uses XGrammar Structural Tag to compose schema-constrained JSONL output.

use_single_sequence pydantic-field

Whether to use a regex that matches exactly one sequence or record if max_sequences_per_example is 1.

GenerateParameters pydantic-model

Bases: Parameters, BaseModel

Configuration parameters for synthetic data generation.

These parameters control how synthetic data is generated after the model is trained. They affect the quality, diversity, and validity of the generated synthetic records.

Fields:

• num_records (int)
• temperature (float)
• repetition_penalty (float)
• top_p (float)
• patience (int)
• invalid_fraction_threshold (float)
• structured_generation (StructuredGenerationParameters)
• enforce_timeseries_fidelity (bool)
• validation (ValidationParameters)
• attention_backend (str | None)

Validators:

• _migrate_legacy_structured_generation_fields

num_records pydantic-field

Number of records to generate.

temperature pydantic-field

Sampling temperature for controlling randomness (higher = more random).

repetition_penalty pydantic-field

The value used to control the likelihood of the model repeating the same token. Must be > 0.

top_p pydantic-field

Nucleus sampling probability for token selection. Must be in (0, 1].

patience pydantic-field

Number of consecutive generations where the invalid_fraction_threshold is reached before stopping generation. Must be >= 1.

invalid_fraction_threshold pydantic-field

The fraction of invalid records that will stop generation after the patience limit is reached. Must be in [0, 1].

structured_generation pydantic-field

Structured generation parameters controlling schema-constrained output.

enforce_timeseries_fidelity pydantic-field

Enforce time-series fidelity by enforcing order, intervals, start and end times of the records.

validation pydantic-field

Validation parameters controlling validation logic and automatic fixes when parsing LLM output and converting to tabular data.

attention_backend pydantic-field

The attention backend for the vLLM engine. Common values: 'FLASHINFER', 'FLASH_ATTN', 'TRITON_ATTN', 'FLEX_ATTENTION'. If None or 'auto', vLLM will auto-select the best available backend.

use_structured_generation property writable

Deprecated flat alias for structured_generation.enabled.

structured_generation_backend property writable

Deprecated flat alias for structured_generation.backend.

structured_generation_schema_method property writable

Deprecated flat alias for structured_generation.schema_method.

structured_generation_use_single_sequence property writable

Deprecated flat alias for structured_generation.use_single_sequence.

resolve_structured_generation_schema_method(schema_method, backend)

Resolve auto schema method from the configured structured-output backend.

auto picks structural_tag on xgrammar-capable backends and regex elsewhere, preserving legacy behavior for outlines/guidance configs that omit an explicit schema method.

Source code in src/nemo_safe_synthesizer/config/generate.py

def resolve_structured_generation_schema_method(
    schema_method: StructuredGenerationSchemaMethod,
    backend: StructuredGenerationBackend | str,
) -> ResolvedStructuredGenerationSchemaMethod:
    """Resolve ``auto`` schema method from the configured structured-output backend.

    ``auto`` picks ``structural_tag`` on xgrammar-capable backends and ``regex``
    elsewhere, preserving legacy behavior for outlines/guidance configs that omit
    an explicit schema method.
    """
    if schema_method != "auto":
        return schema_method
    if backend in STRUCTURAL_TAG_COMPATIBLE_BACKENDS:
        return "structural_tag"
    return "regex"

structural_tag_backend_error_message(backend)

Return an error message when backend cannot serve structural_tag.

vLLM only supports XGrammar Structural Tag constraints when the guided decoding backend is xgrammar or auto (which selects xgrammar for this schema method).

Source code in src/nemo_safe_synthesizer/config/generate.py

def structural_tag_backend_error_message(backend: str) -> str | None:
    """Return an error message when *backend* cannot serve ``structural_tag``.

    vLLM only supports XGrammar Structural Tag constraints when the guided
    decoding backend is ``xgrammar`` or ``auto`` (which selects xgrammar for
    this schema method).
    """
    if backend in STRUCTURAL_TAG_COMPATIBLE_BACKENDS:
        return None
    return (
        "Invalid structured generation configuration: "
        "`schema_method='structural_tag'` requires "
        f"`backend` to be 'xgrammar' or 'auto', got {backend!r}."
    )