File size: 5,123 Bytes
90d08d9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 | #!/usr/bin/env python3
import unittest
from evaluation.run_benchmark import (
dataset_sha256,
parse_decisions,
score_case,
summarize,
)
class ScoreCaseTest(unittest.TestCase):
def test_points_rubric_normalizes_and_applies_penalty(self):
case = {
"grading_method": "points_rubric",
"rubric_items": [
{"item_id": "a", "points": 2},
{"item_id": "b", "points": 1},
{"item_id": "c", "points": -1},
],
}
score, components = score_case(case, {"a": True, "b": False, "c": True})
self.assertAlmostEqual(score, 1 / 3)
self.assertEqual(components["matched_signed_points"], 1)
self.assertEqual(components["available_positive_points"], 3)
def test_points_rubric_clamps_below_zero(self):
case = {
"grading_method": "points_rubric",
"rubric_items": [
{"item_id": "a", "points": 1},
{"item_id": "b", "points": -2},
],
}
score, _ = score_case(case, {"a": False, "b": True})
self.assertEqual(score, 0)
def test_f1_weighted_rubric(self):
case = {
"grading_method": "f1_weighted_rubric",
"rubric_items": [
{"item_id": "a", "points": 3},
{"item_id": "b", "points": 1},
{"item_id": "c", "points": -2},
{"item_id": "d", "points": 0},
],
}
score, components = score_case(
case, {"a": True, "b": False, "c": True, "d": True}
)
# +3/+1/−2/0 → classes 9/7/2/5 → weights 72/3/24/1.
# Recall 72/75; precision 72/97.
recall = 72 / 75
precision = 72 / 97
expected = 2 * precision * recall / (precision + recall)
self.assertAlmostEqual(score, round(expected, 4))
self.assertEqual(components["severe_rate"], 0.0)
self.assertAlmostEqual(components["recall_weighted"], round(recall, 4))
self.assertAlmostEqual(components["precision_weighted"], round(precision, 4))
def test_f1_weighted_rubric_severe_omission_is_reported_not_gated(self):
case = {
"grading_method": "f1_weighted_rubric",
"rubric_items": [
{"item_id": "a", "points": 3},
{"item_id": "b", "points": 2},
],
}
score, components = score_case(case, {"a": False, "b": True})
# +3/+2 → classes 9/8 → weights 72/24. Missing "a" is a severe
# omission: it is reported, but the weighted F1 stands on its own.
recall = 24 / 96
precision = 1.0
expected = 2 * precision * recall / (precision + recall)
self.assertAlmostEqual(score, round(expected, 4))
self.assertEqual(components["severe_rate"], 1.0)
def test_f1_weighted_rubric_severe_commission_is_reported_not_gated(self):
case = {
"grading_method": "f1_weighted_rubric",
"rubric_items": [
{"item_id": "a", "points": 2},
{"item_id": "b", "points": -3},
],
}
score, components = score_case(case, {"a": True, "b": True})
# +2/−3 → classes 8/1 → weights 24/72. Asserting "b" is a severe
# commission: it lands in the precision denominator, not a gate.
recall = 1.0
precision = 24 / 96
expected = 2 * precision * recall / (precision + recall)
self.assertAlmostEqual(score, round(expected, 4))
self.assertEqual(components["severe_rate"], 1.0)
def test_judge_decisions_must_cover_exact_item_ids(self):
case = {
"rubric_items": [
{"item_id": "a"},
{"item_id": "b"},
]
}
with self.assertRaises(ValueError):
parse_decisions(
case,
'{"decisions":[{"item_id":"a","matched":true}]}',
)
def test_dataset_digest_uses_canonical_case_content(self):
first = [{"case_id": "a", "prompt": "test", "rubric_items": []}]
reordered = [{"rubric_items": [], "prompt": "test", "case_id": "a"}]
self.assertEqual(dataset_sha256(first), dataset_sha256(reordered))
def test_summary_includes_reproducibility_metadata(self):
rows = [
{
"case_id": "a",
"benchmark": "example",
"score": 0.5,
"answer_model": "answer",
"judge_model": "judge",
"answer_temperature": 0,
"judge_temperature": None,
"started_at": "2026-01-01T00:00:00+00:00",
"finished_at": "2026-01-01T00:01:00+00:00",
}
]
summary = summarize(rows, 10, 42, {"a"}, "digest")
self.assertEqual(summary["dataset_sha256"], "digest")
self.assertEqual(summary["answer_temperature"], 0)
self.assertEqual(summary["judge_passes_per_answer"], 1)
self.assertEqual(summary["errors"], 0)
if __name__ == "__main__":
unittest.main()
|