Skip to main content

hackagent.datasets.providers.huggingface

HuggingFace dataset provider for loading goals from HuggingFace Hub.

HuggingFaceDatasetProvider Objects​

class HuggingFaceDatasetProvider(DatasetProvider)

Dataset provider for HuggingFace Hub datasets.

This provider loads datasets from HuggingFace Hub and extracts goal strings from specified fields. It supports filtering, splitting, and limiting samples.

Example:

provider = HuggingFaceDatasetProvider({

  • "path" - "ai-safety-institute/AgentHarm",
  • "name" - "harmful",
  • "goal_field" - "prompt",
  • "split" - "test_public", }) goals = provider.load_goals(limit=100)

__init__​

def __init__(config: Dict[str, Any])

Initialize the HuggingFace dataset provider.

Arguments:

  • config - Configuration dictionary with keys:
    • path (str): HuggingFace dataset path (e.g., "ai-safety-institute/AgentHarm")
    • goal_field (str): Field name containing the goal/prompt text
    • split (str, optional): Dataset split to use (default: "test")
    • name (str, optional): Dataset configuration name
    • fallback_fields (list, optional): Alternative fields if goal_field not found
    • trust_remote_code (bool, optional): Whether to trust remote code (default: False)

load_goals​

def load_goals(limit: Optional[int] = None,
               shuffle: bool = False,
               seed: Optional[int] = None,
               **kwargs) -> List[Any]

Load goals from the HuggingFace dataset.

Arguments:

  • limit - Maximum number of goals to return.
  • shuffle - Whether to shuffle the dataset before selecting.
  • seed - Random seed for shuffling.
  • **kwargs - Additional arguments (unused).

Returns:

List of goal strings.

get_extra_data​

def get_extra_data() -> List[Dict[str, Any]]

Return extra fields extracted during the last load_goals call.

get_metadata​

def get_metadata() -> Dict[str, Any]

Return metadata about the loaded dataset.