File size: 11,036 Bytes
09d4173 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 | """Check JevBench conversion, reference alignment and frozen source integrity."""
import copy
import json
import tempfile
import unittest
from dataclasses import replace
from pathlib import Path
from unittest.mock import patch
import httpx
from jev_adapter.benchmarks.data import sha256
from jev_adapter.benchmarks.jevbench import (
JEVBENCH_COMMIT,
JEVBENCH_SUITES,
JevBenchSpec,
normalize_jevbench_record,
parse_jevbench,
prepare_jevbench,
)
def fixture(kind="choice", name="sample"):
questions = {
"choice": {
"type": "choice",
"instructions": "Choose a queue.",
"criteria": {"support": "Support", "billing": "Billing"},
},
"noul": {
"type": "noul",
"instructions": "Is a refund requested?",
"criteria": {"true": "Refund requested", "false": "No refund request"},
},
"score": {
"type": "score",
"instructions": "How urgent?",
"criteria": ["low", "medium", "high"],
},
}
return {
"id": name,
"family": "routing",
"group": None,
"split": "public",
"state": {"message": "Please refund my bill."},
"question": questions[kind],
"labels": {
"choice": ["billing", "support"],
"noul": ["no", "yes"],
"score": ["0", "1", "2"],
}[kind],
"expected": {"choice": "billing", "noul": "yes", "score": 2}[kind],
"provenance": {
"source": "Unit test fixture",
"license": "MIT",
"rationale": "SECRET_RATIONALE",
"surface_answer": "SECRET_SURFACE_ANSWER",
"label_basis": "Authored rubric",
},
}
def source_fixture(records):
payload = "".join(json.dumps(row) + "\n" for row in records).encode()
notices = {"LICENSE": b"Fixture MIT notice", "THIRD-PARTY.md": b"Fixture notice"}
spec = JevBenchSpec(
"datasets/public/fixture.jsonl", sha256(payload), len(records), "Fixture"
)
return (
spec,
{spec.path: payload, **notices},
{path: sha256(content) for path, content in notices.items()},
)
class TestJevBenchNormalization(unittest.TestCase):
def test_native_question_and_canonical_label_orders_preserved_without_gold(self):
raw = fixture()
raw["provenance"]["gold_probs"] = {"support": 0.25, "billing": 0.75}
original = copy.deepcopy(raw)
row = normalize_jevbench_record(raw, "jevbench-hard")
self.assertEqual(raw, original)
self.assertEqual(
row["record"],
{"state": raw["state"], "questions": {"decision": raw["question"]}},
)
expected = row["expected"]["decision"]
self.assertEqual(expected["labels"], ["billing", "support"])
self.assertEqual(
list(row["record"]["questions"]["decision"]["criteria"]),
["support", "billing"],
)
self.assertEqual(expected["target"], [1, 0])
self.assertEqual(expected["reference_probs"], [0.75, 0.25])
request_text = json.dumps(row["record"])
for secret in (
"SECRET_",
"gold_probs",
"expected",
"provenance",
"label_basis",
):
self.assertNotIn(secret, request_text)
self.assertEqual(row["metadata"]["provenance"], raw["provenance"])
self.assertEqual(row["metadata"]["group_id"], raw["id"])
self.assertEqual(row["split"], "public")
def test_noul_reference_mapping_and_score_hard_label(self):
raw = fixture("noul")
raw["provenance"]["gold_probs"] = {"yes": 0.8, "no": 0.2}
noul = normalize_jevbench_record(raw, "jevbench-hard")["expected"]["decision"]
self.assertEqual(noul["labels"], ["false", "true"])
self.assertEqual(noul["target"], [0, 1])
self.assertEqual(noul["reference_probs"], [0.2, 0.8])
score = normalize_jevbench_record(fixture("score"), "jevbench-original")
self.assertEqual(score["expected"]["decision"]["label"], 2)
self.assertNotIn("reference_probs", score["expected"]["decision"])
def test_paraphrase_group_preserved_and_null_groups_independent(self):
first, second = fixture(name="one"), fixture(name="two")
a = normalize_jevbench_record(first, "jevbench-easy")
b = normalize_jevbench_record(second, "jevbench-easy")
self.assertNotEqual(a["metadata"]["group_id"], b["metadata"]["group_id"])
first["group"] = second["group"] = "pair"
self.assertEqual(
normalize_jevbench_record(first, "jevbench-original")["metadata"][
"group_id"
],
normalize_jevbench_record(second, "jevbench-original")["metadata"][
"group_id"
],
)
def test_bad_labels_private_rows_and_malformed_references_rejected(self):
changes = [
{"split": "private"},
{"labels": ["billing", "absent"]},
{"expected": "absent"},
{"provenance": {"exclude_reason": "ambiguous"}},
]
for change in changes:
with self.subTest(change=change), self.assertRaises(ValueError):
normalize_jevbench_record({**fixture(), **change}, "jevbench-hard")
for probs in (
{"billing": 0.5},
{"billing": 0.8, "support": 0.3},
{"billing": float("nan"), "support": 0.3},
{"billing": True, "support": 0.0},
):
raw = fixture()
raw["provenance"]["gold_probs"] = probs
with self.subTest(probs=probs), self.assertRaises(ValueError):
normalize_jevbench_record(raw, "jevbench-hard")
for kind, gold in (("score", True), ("score", 3), ("noul", True)):
raw = fixture(kind)
raw["expected"] = gold
with self.subTest(kind=kind, gold=gold), self.assertRaises(ValueError):
normalize_jevbench_record(raw, "jevbench-hard")
def test_duplicate_and_blank_rows_rejected(self):
row = json.dumps(fixture()) + "\n"
with self.assertRaisesRegex(ValueError, "duplicate"):
parse_jevbench((row * 2).encode(), "jevbench-original")
with self.assertRaisesRegex(ValueError, "blank"):
parse_jevbench((row + "\n").encode(), "jevbench-original")
class TestJevBenchPreparation(unittest.TestCase):
def test_pinned_download_preserves_notices_and_is_idempotent(self):
spec, sources, notices = source_fixture(
[fixture(name="one"), fixture(name="two")]
)
requests = []
def transport(request):
requests.append(request.url.path)
prefix = f"/fstandhartinger/jevbench/{JEVBENCH_COMMIT}/"
self.assertTrue(request.url.path.startswith(prefix))
return httpx.Response(
200, content=sources[request.url.path.removeprefix(prefix)]
)
with (
tempfile.TemporaryDirectory() as tmp,
patch.dict(JEVBENCH_SUITES, {"jevbench-fixture": spec}),
patch("jev_adapter.benchmarks.jevbench.JEVBENCH_NOTICES", notices),
httpx.Client(transport=httpx.MockTransport(transport)) as client,
):
output = Path(tmp)
result = prepare_jevbench(
"jevbench-fixture", output, client=client, limit=1
)
self.assertEqual(
result,
prepare_jevbench("jevbench-fixture", output, client=client, limit=1),
)
self.assertEqual(result["full_partition"]["questions"], 2)
self.assertEqual(result["selected"]["questions"], 1)
self.assertFalse(result["selection"]["is_full_partition"])
directory = output / "jevbench-fixture"
self.assertEqual(
result["data_sha256"], sha256((directory / "public.jsonl").read_bytes())
)
for name in notices:
self.assertEqual((directory / name).read_bytes(), sources[name])
self.assertTrue(
all("private" not in path and "train" not in path for path in requests)
)
with self.assertRaises(FileExistsError):
prepare_jevbench("jevbench-fixture", output, client=client)
self.assertEqual(
result["data_sha256"], sha256((directory / "public.jsonl").read_bytes())
)
def test_bad_data_notice_or_count_fails_before_writing(self):
spec, sources, notices = source_fixture([fixture()])
for corrupt in (spec.path, "LICENSE", "THIRD-PARTY.md", "count"):
with (
self.subTest(corrupt=corrupt),
tempfile.TemporaryDirectory() as tmp,
patch.dict(
JEVBENCH_SUITES,
{
"jevbench-fixture": replace(spec, records=2)
if corrupt == "count"
else spec
},
),
patch("jev_adapter.benchmarks.jevbench.JEVBENCH_NOTICES", notices),
):
root, output = Path(tmp) / "source", Path(tmp) / "out"
for name, content in sources.items():
path = root / name
path.parent.mkdir(parents=True, exist_ok=True)
path.write_bytes(content + (b" " if name == corrupt else b""))
with self.assertRaisesRegex(
ValueError, "SHA256 mismatch|count mismatch"
):
prepare_jevbench(
"jevbench-fixture", output, source_root=root, limit=1
)
self.assertFalse(output.exists())
def test_changed_notice_is_not_partially_overwritten(self):
spec, sources, notices = source_fixture([fixture()])
with (
tempfile.TemporaryDirectory() as tmp,
patch.dict(JEVBENCH_SUITES, {"jevbench-fixture": spec}),
patch("jev_adapter.benchmarks.jevbench.JEVBENCH_NOTICES", notices),
):
root, output = Path(tmp) / "source", Path(tmp) / "out"
for name, content in sources.items():
path = root / name
path.parent.mkdir(parents=True, exist_ok=True)
path.write_bytes(content)
directory = output / "jevbench-fixture"
directory.mkdir(parents=True)
(directory / "LICENSE").write_bytes(b"Existing different license")
with self.assertRaises(FileExistsError):
prepare_jevbench("jevbench-fixture", output, source_root=root)
self.assertFalse((directory / "public.jsonl").exists())
self.assertEqual(
(directory / "LICENSE").read_bytes(), b"Existing different license"
)
if __name__ == "__main__":
unittest.main()
|