Skip to content

DeepTeam is a framework to red team LLMs and AI agents.

.

A-Vamshidf705c782701
diff --git a/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py b/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
index 0458b9aa..a4742fd9 100644
--- a/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
@@ -109,7 +109,7 @@ class TreeJailbreaking(BaseMultiTurnAttack):
scored = []
for probe in probes:
score = self._score(progression, probe)
- progression.score_probe(probe, score)
+ progression.score_probe(probe, self._unit_score(score))
scored.append((probe, score))
best, best_score = max(scored, key=lambda pair: pair[1])
progression.commit(best)
@@ -165,7 +165,7 @@ class TreeJailbreaking(BaseMultiTurnAttack):
]
)
for probe, score in zip(probes, scores):
- progression.score_probe(probe, score)
+ progression.score_probe(probe, self._unit_score(score))
best, best_score = max(
zip(probes, scores), key=lambda pair: pair[1]
)
@@ -296,6 +296,10 @@ class TreeJailbreaking(BaseMultiTurnAttack):
return res.prompt, topic.on_topic
+ @staticmethod
+ def _unit_score(rating: float) -> float:
+ return max(0.0, min(1.0, (rating - 1) / 9))
+
def _score(self, progression: Progression, probe: Probe) -> float:
rating: Rating = generate(
JailBreakingTemplate.linear_judge(