jyapayne/bypass-prompt-guard-2
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Fix running

Browse source
This commit is contained in:
Joey Yakimowich-Payne 2025-05-12 16:54:01 -06:00
commit 155d8cf80a
3 changed files with 30 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

8
hacking.py
View file

@ -330,8 +330,8 @@ def main():
if best_prefix is not None: if best_prefix is not None:
best_prefix = minimize_tokens( best_prefix = minimize_tokens(
model, tokenizer, injection_text, best_prefix, text, model, tokenizer, injection_text, best_prefix, text,
benign_class_idx, min_benign_confidence, device=device, benign_class_idx, device=device,
target_tokens=1, min_acceptable_benign=min_acceptable_benign, min_acceptable_benign=min_acceptable_benign,
) )
else: else:
print("\n===== Did not find a high confidence benign classification =====") print("\n===== Did not find a high confidence benign classification =====")
@ -341,8 +341,8 @@ def main():
# Still try to minimize tokens # Still try to minimize tokens
best_prefix = minimize_tokens( best_prefix = minimize_tokens(
model, tokenizer, injection_text, best_prefix, text, model, tokenizer, injection_text, best_prefix, text,
benign_class_idx, best_score * 0.95, target_tokens=1, benign_class_idx, device=device,
min_acceptable_benign=min_acceptable_benign, device=device, min_acceptable_benign=min_acceptable_benign,
) )
# Use the best prefix found across all runs # Use the best prefix found across all runs

5
utils.py
View file

@ -410,10 +410,8 @@ def minimize_tokens(
adv_prefix: str, adv_prefix: str,
text: str, text: str,
benign_class_idx: int, benign_class_idx: int,
min_benign_confidence: float,
device: torch.device, device: torch.device,
min_acceptable_benign: float = 0.6, min_acceptable_benign: float = 0.6,
token_length_weight: float = 0.3 # Weight for prioritizing removal of short tokens
) -> str: ) -> str:
""" """
Minimize tokens using only token contribution analysis (ablation study). Minimize tokens using only token contribution analysis (ablation study).
@ -429,9 +427,8 @@ def minimize_tokens(
# Use only token ablation approach - systematically remove tokens that contribute least # Use only token ablation approach - systematically remove tokens that contribute least
ablation_prefix: str = analyze_token_contributions( ablation_prefix: str = analyze_token_contributions(
model, tokenizer, injection_text, adv_prefix, text, model, tokenizer, injection_text, adv_prefix, text,
benign_class_idx, min_benign_confidence=min_benign_confidence, benign_class_idx,
device=device, min_acceptable_benign=min_acceptable_benign, device=device, min_acceptable_benign=min_acceptable_benign,
token_length_weight=token_length_weight # Pass through the token length weight
) )
# Report final token count # Report final token count

25
words.py
View file

@ -465,6 +465,31 @@ words3 = list(set(words3))
words4 = [ words4 = [
"ocular",
"spell",
"License",
"paper",
"updated",
"others",
"boards",
"better",
"separate",
"категор",
"许可证",
"Grow",
"All",
"Simon",
"contest",
"all",
"Categories",
"bean",
"elastic",
"variable",
"generator",
"Arizona",
"Itemid",
"recommend",
"better",
"previous", "previous",
"Camp", "Camp",
"Megan", "Megan",