Models

The models module defines configuration objects for model-based generation. ModelProvider, specifies connection and authentication details for custom providers. ModelConfig encapsulates model details including the model alias, identifier, and inference parameters. Inference Parameters controls model behavior through settings like temperature, top_p, and max_tokens, with support for both fixed values and distribution-based sampling. The module includes ImageContext for providing image inputs to multimodal models.

For more information on how they are used, see below:

• Model Providers
• Model Configs
• Image Context

Classes:

Name	Description
BaseInferenceParams	Base configuration for inference parameters.
ChatCompletionInferenceParams	Configuration for LLM inference parameters.
DistributionType	Types of distributions for sampling inference parameters.
EmbeddingInferenceParams	Configuration for embedding generation parameters.
ImageContext	Configuration for providing image context to multimodal models.
ImageFormat	Supported image formats for image modality.
ManualDistribution	Manual (discrete) distribution for sampling inference parameters.
ManualDistributionParams	Parameters for manual distribution sampling.
Modality	Supported modality types for multimodal model data.
ModalityDataType	Data type formats for multimodal data.
ModelConfig	Configuration for a model used for generation.
ModelProvider	Configuration for a custom model provider.
UniformDistribution	Uniform distribution for sampling inference parameters.
UniformDistributionParams	Parameters for uniform distribution sampling.

BaseInferenceParams

Bases: ConfigBase, ABC

Base configuration for inference parameters.

Attributes:

Name	Type	Description
generation_type	GenerationType	Type of generation (chat-completion or embedding). Acts as discriminator.
max_parallel_requests	int	Maximum number of parallel requests to the model API.
timeout	int | None	Timeout in seconds for each request.
extra_body	dict[str, Any] | None	Additional parameters to pass to the model API.

Methods:

Name	Description
format_for_display	Format inference parameters for display.

generate_kwargs property

Get the generate kwargs for the inference parameters.

Returns:

Type	Description
dict[str, Any]	A dictionary of the generate kwargs.

format_for_display()

Format inference parameters for display.

Returns:

Type	Description
str	Formatted string of inference parameters

Source code in src/data_designer/config/models.py

def format_for_display(self) -> str:
    """Format inference parameters for display.

    Returns:
        Formatted string of inference parameters
    """
    params_dict = self.model_dump(exclude_none=True, mode="json")

    if not params_dict:
        return "(none)"

    parts = []
    for key, value in params_dict.items():
        formatted_value = self._format_value(key, value)
        parts.append(f"{key}={formatted_value}")
    return ", ".join(parts)

ChatCompletionInferenceParams

Bases: BaseInferenceParams

Configuration for LLM inference parameters.

Attributes:

Name	Type	Description
generation_type	Literal[CHAT_COMPLETION]	Type of generation, always "chat-completion" for this class.
temperature	float | DistributionT | None	Sampling temperature (0.0-2.0). Can be a fixed value or a distribution for dynamic sampling.
top_p	float | DistributionT | None	Nucleus sampling probability (0.0-1.0). Can be a fixed value or a distribution for dynamic sampling.
max_tokens	int | None	Maximum number of tokens to generate in the response.

DistributionType

Bases: str, Enum

Types of distributions for sampling inference parameters.

EmbeddingInferenceParams

Bases: BaseInferenceParams

Configuration for embedding generation parameters.

Attributes:

Name	Type	Description
generation_type	Literal[EMBEDDING]	Type of generation, always "embedding" for this class.
encoding_format	Literal['float', 'base64']	Format of the embedding encoding ("float" or "base64").
dimensions	int | None	Number of dimensions for the embedding.

ImageContext

Bases: ModalityContext

Configuration for providing image context to multimodal models.

Attributes:

Name	Type	Description
modality	Modality	The modality type (always "image").
column_name	str	Name of the column containing image data.
data_type	ModalityDataType	Format of the image data ("url" or "base64").
image_format	ImageFormat | None	Image format (required for base64 data).

Methods:

Name	Description
get_context	Get the context for the image modality.

get_context(record)

Get the context for the image modality.

Parameters:

Name	Type	Description	Default
record	dict	The record containing the image data.	required

Returns:

Type	Description
dict[str, Any]	The context for the image modality.

Source code in src/data_designer/config/models.py

def get_context(self, record: dict) -> dict[str, Any]:
    """Get the context for the image modality.

    Args:
        record: The record containing the image data.

    Returns:
        The context for the image modality.
    """
    context = dict(type="image_url")
    context_value = record[self.column_name]
    if self.data_type == ModalityDataType.URL:
        context["image_url"] = context_value
    else:
        context["image_url"] = {
            "url": f"data:image/{self.image_format.value};base64,{context_value}",
            "format": self.image_format.value,
        }
    return context

ImageFormat

Bases: str, Enum

Supported image formats for image modality.

ManualDistribution

Bases: Distribution[ManualDistributionParams]

Manual (discrete) distribution for sampling inference parameters.

Samples from a discrete set of values with optional weights. Useful for testing specific values or creating custom probability distributions for temperature or top_p.

Attributes:

Name	Type	Description
distribution_type	DistributionType | None	Type of distribution ("manual").
params	ManualDistributionParams	Distribution parameters (values, weights).

Methods:

Name	Description
sample	Sample a value from the manual distribution.

sample()

Sample a value from the manual distribution.

Returns:

Type	Description
float	A float value sampled from the manual distribution.

Source code in src/data_designer/config/models.py

def sample(self) -> float:
    """Sample a value from the manual distribution.

    Returns:
        A float value sampled from the manual distribution.
    """
    return float(np.random.choice(self.params.values, p=self.params.weights))

ManualDistributionParams

Bases: ConfigBase

Parameters for manual distribution sampling.

Attributes:

Name	Type	Description
values	list[float]	List of possible values to sample from.
weights	list[float] | None	Optional list of weights for each value. If not provided, all values have equal probability.

Modality

Bases: str, Enum

Supported modality types for multimodal model data.

ModalityDataType

Bases: str, Enum

Data type formats for multimodal data.

ModelConfig

Bases: ConfigBase

Configuration for a model used for generation.

Attributes:

Name	Type	Description
alias	str	User-defined alias to reference in column configurations.
model	str	Model identifier (e.g., from build.nvidia.com or other providers).
inference_parameters	InferenceParamsT	Inference parameters for the model (temperature, top_p, max_tokens, etc.). The generation_type is determined by the type of inference_parameters.
provider	str | None	Optional model provider name if using custom providers.

generation_type property

Get the generation type from the inference parameters.

ModelProvider

Bases: ConfigBase

Configuration for a custom model provider.

Attributes:

Name	Type	Description
name	str	Name of the model provider.
endpoint	str	API endpoint URL for the provider.
provider_type	str	Provider type (default: "openai"). Determines the API format to use.
api_key	str | None	Optional API key for authentication.
extra_body	dict[str, Any] | None	Additional parameters to pass in API requests.
extra_headers	dict[str, str] | None	Additional headers to pass in API requests.

UniformDistribution

Bases: Distribution[UniformDistributionParams]

Uniform distribution for sampling inference parameters.

Samples values uniformly between low and high bounds. Useful for exploring a continuous range of values for temperature or top_p.

Attributes:

Name	Type	Description
distribution_type	DistributionType | None	Type of distribution ("uniform").
params	UniformDistributionParams	Distribution parameters (low, high).

Methods:

Name	Description
sample	Sample a value from the uniform distribution.

sample()

Sample a value from the uniform distribution.

Returns:

Type	Description
float	A float value sampled from the uniform distribution.

Source code in src/data_designer/config/models.py

def sample(self) -> float:
    """Sample a value from the uniform distribution.

    Returns:
        A float value sampled from the uniform distribution.
    """
    return float(np.random.uniform(low=self.params.low, high=self.params.high, size=1)[0])

UniformDistributionParams

Bases: ConfigBase

Parameters for uniform distribution sampling.

Attributes:

Name	Type	Description
low	float	Lower bound (inclusive).
high	float	Upper bound (exclusive).