api_reference.md

API Reference

load_wildguard

def load_wildguard(
    use_vllm: bool = True,
    device: str = 'cuda',
    ephemeral_model: bool = True,
    batch_size: int | None = None,
) -> WildGuard:

Loads and returns a WildGuard model for classification.

Parameters

• use_vllm (bool, optional): Whether to use a VLLM model for classification. If False, uses a HuggingFace model. Default is True.
• device (str, optional): The device to run the HuggingFace model on. Ignored if using VLLM. Default is 'cuda'.
• ephemeral_model (bool, optional): Whether to remove the model from the device and free GPU memory after calling classify(). Default is True.
• batch_size (int | None, optional): The batch size for classification. If None, uses default values (16 for HF, -1 for VLLM).

Returns

• WildGuard: An instance of the WildGuard class.

WildGuard.classify

def classify(
    self,
    items: list[dict[str, str]],
    save_func: Callable[[list[dict[str, Any]]], None] | None = None
) -> list[dict[str, Any]]:

Classifies a list of items for safety.

Parameters

• items (list[dict[str, str]]): A list of dictionaries, each containing a 'prompt' key and optionally a 'response' key.
• save_func (Callable[[list[dict[str, Any]]], None] | None, optional): A function to save intermediate results. Default is None.

Returns

• list[dict[str, Any]]: A list of dictionaries containing classification results. Each item contains the following fields:
  • 'prompt_harmfulness' (str): either 'harmful' or 'unharmful'
  • 'response_harmfulness' (str | None): either 'harmful' or 'unharmful'. If a response was not provided in the input, it is None.
  • 'response_refusal' (str | None): either 'refusal' or 'compliance'. If a response was not provided in the input, it is None.
  • 'is_parsing_error' (bool): True if parsing the model output failed. If this is True, the rest of the results may be invalid.