Attack goal function details

[ary4n99]

Hey, what's the reason behind using the difference in scores for the attack goal function instead of the PDR?

promptbench/promptbench/prompt_attack/goal_function.py

Lines 276 to 279 in 3daaf2d

	def _get_score(self, model_output, _):
	score = self.ground_truth_output - model_output
	return score

Also, why has maximizable=True not been passed to the goal function?

promptbench/promptbench/prompt_attack/goal_function.py

Lines 242 to 248 in 3daaf2d

	class AdvPromptGoalFunction(GoalFunction):
	"""A goal function defined on a model that outputs a probability for some
	number of classes."""
	def __init__(self, model, dataset, eval_func, query_budget, *args, target_max_acc=0, verbose=True, **kwargs):
	super().__init__(*args, **kwargs)

[Immortalise]

1. Using PDR is also beneficial, I think it has a similar impact as the score differences.
2. In the adversarial prompts goal function, the maximizable argument is redundant.

[ary4n99]

Thank you for clarifying!