hackagent.datasets.registry

Registry and factory functions for dataset providers.

This module provides the main entry point for loading goals from various sources.

register_provider

def register_provider(name: str,
                      provider_class: Type[DatasetProvider]) -> None

Register a dataset provider.

Arguments:

• name - The provider name.
• provider_class - The provider class.

get_provider

def get_provider(name: str, config: Dict[str, Any]) -> DatasetProvider

Get a dataset provider instance by name.

Arguments:

• name - The provider name (e.g., "huggingface", "file").
• config - Provider configuration dictionary.

Returns:

Configured DatasetProvider instance.

Raises:

• ValueError - If the provider is not found.

load_goals

def load_goals(preset: Optional[str] = None,
               provider: Optional[str] = None,
               path: Optional[str] = None,
               goal_field: Optional[str] = None,
               split: Optional[str] = None,
               name: Optional[str] = None,
               limit: Optional[int] = None,
               shuffle: bool = False,
               seed: Optional[int] = None,
               **kwargs) -> List[str]

Load attack goals from a dataset source.

This is the main entry point for loading goals. It supports three modes:

1. Preset mode: Use a pre-configured dataset by name
2. Provider mode: Directly specify provider and dataset details
3. Config mode: Pass a full configuration dictionary

Arguments:

• preset - Name of a pre-configured dataset preset (e.g., "agentharm", "strongreject").
• provider - Provider type ("huggingface" or "file").
• path - Dataset path (HuggingFace dataset ID or file path).
• goal_field - Field name containing the goal text.
• split - Dataset split to use (for HuggingFace).
• name - Dataset configuration name (for HuggingFace datasets with multiple configs).
• limit - Maximum number of goals to return.
• shuffle - Whether to shuffle before selecting.
• seed - Random seed for shuffling.
• **kwargs - Additional provider-specific configuration.

Returns:

List of goal strings.

Examples:

Using a preset

goals = load_goals(preset="agentharm", limit=50)

Using HuggingFace directly

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

Using a local file

goals = load_goals( provider="file", path="./my_goals.json", goal_field="objective" )

Raises:

• provider0 - If neither preset nor provider is specified.

load_goals_from_config

def load_goals_from_config(config: Dict[str, Any]) -> List[str]

Load goals from a configuration dictionary.

This function is designed to be called from the AttackOrchestrator when a 'dataset' key is present in the attack configuration.

Arguments:

• config - Dataset configuration dictionary with keys:
  • preset (str, optional): Preset name
  • provider (str, optional): Provider type
  • path (str, optional): Dataset path
  • goal_field (str, optional): Field containing goals
  • split (str, optional): Dataset split
  • name (str, optional): Dataset config name
  • limit (int, optional): Max goals to load
  • shuffle (bool, optional): Shuffle before selecting
  • seed (int, optional): Random seed

Returns:

List of goal strings.

Example config: {

• "preset" - "agentharm",

• "limit" - 100,

• "shuffle" - True }

  Or:

  {

• "provider" - "huggingface",

• "path" - "ai-safety-institute/AgentHarm",

• "goal_field" - "prompt",

• "split" - "test_public",

• "limit" - 50 }