Spaces:
Sleeping
Sleeping
Commit ·
843ee19
1
Parent(s): a5aa26d
test(graders): add medium_grader tests — diversity scores higher than uniform strategies
Browse files- tests/test_graders.py +16 -7
tests/test_graders.py
CHANGED
|
@@ -1,17 +1,26 @@
|
|
| 1 |
from graders.easy_grader import grade_easy
|
|
|
|
| 2 |
|
| 3 |
def test_easy_grader():
|
| 4 |
-
# Empty history
|
| 5 |
assert grade_easy([]) == 0.0
|
| 6 |
-
|
| 7 |
-
# Successful attack
|
| 8 |
history = [{"attack_success_estimate": 0.9}]
|
| 9 |
assert grade_easy(history) == 0.9
|
| 10 |
-
|
| 11 |
-
# Failed attack
|
| 12 |
history = [{"attack_success_estimate": 0.1}]
|
| 13 |
assert grade_easy(history) == 0.1
|
| 14 |
-
|
| 15 |
-
# Range check
|
| 16 |
history = [{"attack_success_estimate": 1.5}]
|
| 17 |
assert grade_easy(history) == 1.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
from graders.easy_grader import grade_easy
|
| 2 |
+
from graders.medium_grader import grade_medium
|
| 3 |
|
| 4 |
def test_easy_grader():
|
|
|
|
| 5 |
assert grade_easy([]) == 0.0
|
|
|
|
|
|
|
| 6 |
history = [{"attack_success_estimate": 0.9}]
|
| 7 |
assert grade_easy(history) == 0.9
|
|
|
|
|
|
|
| 8 |
history = [{"attack_success_estimate": 0.1}]
|
| 9 |
assert grade_easy(history) == 0.1
|
|
|
|
|
|
|
| 10 |
history = [{"attack_success_estimate": 1.5}]
|
| 11 |
assert grade_easy(history) == 1.0
|
| 12 |
+
|
| 13 |
+
def test_medium_grader_diversity():
|
| 14 |
+
# Uniform strategies
|
| 15 |
+
h1 = [
|
| 16 |
+
{"attack_success_estimate": 0.5, "strategy_type": "roleplay"},
|
| 17 |
+
{"attack_success_estimate": 0.5, "strategy_type": "roleplay"}
|
| 18 |
+
]
|
| 19 |
+
|
| 20 |
+
# Diverse strategies
|
| 21 |
+
h2 = [
|
| 22 |
+
{"attack_success_estimate": 0.5, "strategy_type": "roleplay"},
|
| 23 |
+
{"attack_success_estimate": 0.5, "strategy_type": "injection"}
|
| 24 |
+
]
|
| 25 |
+
|
| 26 |
+
assert grade_medium(h2) > grade_medium(h1)
|