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

Remove more bible

Browse source
This commit is contained in:
Joey Yakimowich-Payne 2025-05-12 14:42:39 -06:00
commit 329f4e668e
2 changed files with 3 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

2
hacking.py
View file

@ -19,8 +19,6 @@ device: torch.device = torch.device("cuda" if cuda_available else "cpu")
# use token from environment variable # use token from environment variable
login(token=os.getenv("HF_TOKEN")) login(token=os.getenv("HF_TOKEN"))
bible_words: List[str] = open("bible.txt").read().split()
alpha: float = 0.3 # weight between benign maximization and loss minimization alpha: float = 0.3 # weight between benign maximization and loss minimization
min_benign_confidence: float = 0.95 # Minimum benign probability to stop (95%) min_benign_confidence: float = 0.95 # Minimum benign probability to stop (95%)
words_to_inject: int = 1 words_to_inject: int = 1

8
utils.py
View file

@ -150,6 +150,8 @@ def find_best_word_to_add(
results = [] results = []
best_combined_score = 0 best_combined_score = 0
best_result_idx = -1 best_result_idx = -1
print(f"Running score analysis for {len(all_candidate_prefixes)} candidate prefixes")
for idx, candidate in enumerate(all_candidate_prefixes): for idx, candidate in enumerate(all_candidate_prefixes):
benign_score = benign_scores[idx] benign_score = benign_scores[idx]
@ -187,7 +189,7 @@ def find_best_word_to_add(
token_count, combined_score token_count, combined_score
) )
print(f"Word '{candidate['word']}' at {candidate['position']}: {benign_score:.4f} (Δ: {improvement:.4f}, tokens: {token_count}, combined: {combined_score:.4f})") #print(f"Word '{candidate['word']}' at {candidate['position']}: {benign_score:.4f} (Δ: {improvement:.4f}, tokens: {token_count}, combined: {combined_score:.4f})")
# Only consider improvements (benign_score > baseline_score) # Only consider improvements (benign_score > baseline_score)
if improvement > 0 and combined_score > best_combined_score: if improvement > 0 and combined_score > best_combined_score:
@ -297,10 +299,8 @@ def analyze_token_contributions(
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
order_template: str = "{injection}{prefix}{text}" # Template for ordering components order_template: str = "{injection}{prefix}{text}" # Template for ordering components
) -> str: ) -> str:
""" """
@ -412,7 +412,6 @@ def minimize_tokens(
benign_class_idx: int, benign_class_idx: int,
min_benign_confidence: float, min_benign_confidence: float,
device: torch.device, device: torch.device,
target_tokens: int = 1,
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 token_length_weight: float = 0.3 # Weight for prioritizing removal of short tokens
) -> str: ) -> str:
@ -543,7 +542,6 @@ def get_combined_score(
text: str, text: str,
candidates: List[str], candidates: List[str],
benign_idx: int, benign_idx: int,
malicious_idx: int,
device: torch.device, device: torch.device,
alpha: float = 0.5, alpha: float = 0.5,
token_penalty_weight: float = 0.1, token_penalty_weight: float = 0.1,