Spaces:
Sleeping
Sleeping
Upload redveil/grader.py with huggingface_hub
Browse files- redveil/grader.py +11 -12
redveil/grader.py
CHANGED
|
@@ -1,14 +1,13 @@
|
|
| 1 |
"""Grader system for RedVeil tasks.
|
| 2 |
|
| 3 |
Each grader checks the environment's internal game state and returns
|
| 4 |
-
a score strictly between 0 and
|
| 5 |
|
| 6 |
Honeypot penalties reduce score. Efficiency bonuses reward clean play.
|
| 7 |
"""
|
| 8 |
|
| 9 |
from .tasks import ALL_TASKS
|
| 10 |
|
| 11 |
-
# Validator requires scores strictly in (0, 1) — never exactly 0.0 or 1.0
|
| 12 |
_SCORE_MIN = 0.01
|
| 13 |
_SCORE_MAX = 0.99
|
| 14 |
|
|
@@ -25,7 +24,7 @@ def grade_task(game_state: dict) -> float:
|
|
| 25 |
game_state: Dictionary from RedVeilEnvironment.get_game_state()
|
| 26 |
|
| 27 |
Returns:
|
| 28 |
-
Score strictly between 0 and
|
| 29 |
"""
|
| 30 |
task_id = game_state.get("task_id")
|
| 31 |
if task_id == "easy_recon":
|
|
@@ -36,7 +35,7 @@ def grade_task(game_state: dict) -> float:
|
|
| 36 |
return grade_hard(game_state)
|
| 37 |
elif task_id == "expert_chain":
|
| 38 |
return grade_expert(game_state)
|
| 39 |
-
return
|
| 40 |
|
| 41 |
|
| 42 |
def _honeypot_penalty(state: dict) -> float:
|
|
@@ -52,7 +51,7 @@ def grade_easy(state: dict) -> float:
|
|
| 52 |
scanned_port -> 0.1
|
| 53 |
identified_1_real -> 0.3
|
| 54 |
identified_2_real -> 0.6
|
| 55 |
-
identified_all_3_clean ->
|
| 56 |
|
| 57 |
Penalty: false positives (identifying noise as real) reduce score.
|
| 58 |
"""
|
|
@@ -77,7 +76,7 @@ def grade_easy(state: dict) -> float:
|
|
| 77 |
return _clamp(score)
|
| 78 |
elif scanned:
|
| 79 |
return _clamp(0.1)
|
| 80 |
-
return
|
| 81 |
|
| 82 |
|
| 83 |
def grade_medium(state: dict) -> float:
|
|
@@ -87,7 +86,7 @@ def grade_medium(state: dict) -> float:
|
|
| 87 |
discovered_endpoint -> 0.15
|
| 88 |
found_real_vuln -> 0.4
|
| 89 |
exploited_vuln -> 0.7
|
| 90 |
-
flagged_honeypot ->
|
| 91 |
|
| 92 |
Penalty for honeypot interactions.
|
| 93 |
"""
|
|
@@ -106,7 +105,7 @@ def grade_medium(state: dict) -> float:
|
|
| 106 |
return _clamp(0.4 - penalty)
|
| 107 |
elif discovered:
|
| 108 |
return _clamp(0.15)
|
| 109 |
-
return
|
| 110 |
|
| 111 |
|
| 112 |
def grade_hard(state: dict) -> float:
|
|
@@ -118,7 +117,7 @@ def grade_hard(state: dict) -> float:
|
|
| 118 |
found_real_vuln -> 0.3
|
| 119 |
exploited_vuln -> 0.55
|
| 120 |
extracted_creds -> 0.75
|
| 121 |
-
admin_login ->
|
| 122 |
|
| 123 |
Penalty for honeypot interactions.
|
| 124 |
"""
|
|
@@ -143,7 +142,7 @@ def grade_hard(state: dict) -> float:
|
|
| 143 |
return _clamp(0.15)
|
| 144 |
elif has_recon:
|
| 145 |
return _clamp(0.05)
|
| 146 |
-
return
|
| 147 |
|
| 148 |
|
| 149 |
def grade_expert(state: dict) -> float:
|
|
@@ -155,7 +154,7 @@ def grade_expert(state: dict) -> float:
|
|
| 155 |
low_priv_access -> 0.25
|
| 156 |
acquired_token -> 0.4
|
| 157 |
extracted_admin_creds -> 0.7
|
| 158 |
-
admin_login ->
|
| 159 |
|
| 160 |
Heavy penalty for honeypot interactions.
|
| 161 |
"""
|
|
@@ -180,4 +179,4 @@ def grade_expert(state: dict) -> float:
|
|
| 180 |
return _clamp(0.12)
|
| 181 |
elif has_recon:
|
| 182 |
return _clamp(0.05)
|
| 183 |
-
return
|
|
|
|
| 1 |
"""Grader system for RedVeil tasks.
|
| 2 |
|
| 3 |
Each grader checks the environment's internal game state and returns
|
| 4 |
+
a score strictly between 0.01 and 0.99 based on milestone progression.
|
| 5 |
|
| 6 |
Honeypot penalties reduce score. Efficiency bonuses reward clean play.
|
| 7 |
"""
|
| 8 |
|
| 9 |
from .tasks import ALL_TASKS
|
| 10 |
|
|
|
|
| 11 |
_SCORE_MIN = 0.01
|
| 12 |
_SCORE_MAX = 0.99
|
| 13 |
|
|
|
|
| 24 |
game_state: Dictionary from RedVeilEnvironment.get_game_state()
|
| 25 |
|
| 26 |
Returns:
|
| 27 |
+
Score strictly between 0.01 and 0.99
|
| 28 |
"""
|
| 29 |
task_id = game_state.get("task_id")
|
| 30 |
if task_id == "easy_recon":
|
|
|
|
| 35 |
return grade_hard(game_state)
|
| 36 |
elif task_id == "expert_chain":
|
| 37 |
return grade_expert(game_state)
|
| 38 |
+
return _SCORE_MIN
|
| 39 |
|
| 40 |
|
| 41 |
def _honeypot_penalty(state: dict) -> float:
|
|
|
|
| 51 |
scanned_port -> 0.1
|
| 52 |
identified_1_real -> 0.3
|
| 53 |
identified_2_real -> 0.6
|
| 54 |
+
identified_all_3_clean -> 0.99
|
| 55 |
|
| 56 |
Penalty: false positives (identifying noise as real) reduce score.
|
| 57 |
"""
|
|
|
|
| 76 |
return _clamp(score)
|
| 77 |
elif scanned:
|
| 78 |
return _clamp(0.1)
|
| 79 |
+
return _SCORE_MIN
|
| 80 |
|
| 81 |
|
| 82 |
def grade_medium(state: dict) -> float:
|
|
|
|
| 86 |
discovered_endpoint -> 0.15
|
| 87 |
found_real_vuln -> 0.4
|
| 88 |
exploited_vuln -> 0.7
|
| 89 |
+
flagged_honeypot -> 0.99
|
| 90 |
|
| 91 |
Penalty for honeypot interactions.
|
| 92 |
"""
|
|
|
|
| 105 |
return _clamp(0.4 - penalty)
|
| 106 |
elif discovered:
|
| 107 |
return _clamp(0.15)
|
| 108 |
+
return _SCORE_MIN
|
| 109 |
|
| 110 |
|
| 111 |
def grade_hard(state: dict) -> float:
|
|
|
|
| 117 |
found_real_vuln -> 0.3
|
| 118 |
exploited_vuln -> 0.55
|
| 119 |
extracted_creds -> 0.75
|
| 120 |
+
admin_login -> 0.99
|
| 121 |
|
| 122 |
Penalty for honeypot interactions.
|
| 123 |
"""
|
|
|
|
| 142 |
return _clamp(0.15)
|
| 143 |
elif has_recon:
|
| 144 |
return _clamp(0.05)
|
| 145 |
+
return _SCORE_MIN
|
| 146 |
|
| 147 |
|
| 148 |
def grade_expert(state: dict) -> float:
|
|
|
|
| 154 |
low_priv_access -> 0.25
|
| 155 |
acquired_token -> 0.4
|
| 156 |
extracted_admin_creds -> 0.7
|
| 157 |
+
admin_login -> 0.99
|
| 158 |
|
| 159 |
Heavy penalty for honeypot interactions.
|
| 160 |
"""
|
|
|
|
| 179 |
return _clamp(0.12)
|
| 180 |
elif has_recon:
|
| 181 |
return _clamp(0.05)
|
| 182 |
+
return _SCORE_MIN
|