Spaces:
Sleeping
Sleeping
Commit ·
333177f
1
Parent(s): c6ae20e
feat(environment): integrate reward_computer call in step() with mock fallback
Browse files- server/environment.py +17 -0
server/environment.py
CHANGED
|
@@ -82,4 +82,21 @@ class RedTeamEnvironment:
|
|
| 82 |
"defense_score": 0.9,
|
| 83 |
"llm_grade": {"overall": 3.0},
|
| 84 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
return None # Placeholder
|
|
|
|
| 82 |
"defense_score": 0.9,
|
| 83 |
"llm_grade": {"overall": 3.0},
|
| 84 |
}
|
| 85 |
+
|
| 86 |
+
# ------ Step 4: Call Person 2's reward computer ------
|
| 87 |
+
if self.reward_computer:
|
| 88 |
+
reward_result = self.reward_computer(
|
| 89 |
+
action = action,
|
| 90 |
+
attack_history = self.attack_history,
|
| 91 |
+
llm_result = llm_result,
|
| 92 |
+
)
|
| 93 |
+
else:
|
| 94 |
+
# Mock — lets you run the server before P2 is ready
|
| 95 |
+
logger.warning("Reward computer not wired — using mock reward")
|
| 96 |
+
reward_result = {
|
| 97 |
+
"total_reward": 0.0,
|
| 98 |
+
"novelty_score": 0.5,
|
| 99 |
+
"feedback": "[Mock] Reward system not connected.",
|
| 100 |
+
"safety_flagged": False,
|
| 101 |
+
}
|
| 102 |
return None # Placeholder
|