blakeslee-research/nucleic-purpose-classifier
3
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge nucleic/fuzzy-dewy-urchin-hpvd into dev

Browse Source
This commit is contained in:
abkslm
2026-08-02 16:38:29 -07:00
parent f89561b6a8
commit 4ad5b65811
3 changed files with 127 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test_rebuild_sol_high.py
+50 -3
View File
@@ -185,7 +185,18 @@ class RebuildSolHighTests(unittest.TestCase):
history = root / "history.jsonl"
purpose_data.write_jsonl(history, [{"prompt": "Add a private cache layer"}])
swe = root / "swe.jsonl"
purpose_data.write_jsonl(swe, [{"prompt": "Diagnose a unique worker crash"}])
duplicate_words = [f"duplicate{index}" for index in range(100)]
first_duplicate = " ".join(duplicate_words)
duplicate_words[50] = "replacement"
second_duplicate = " ".join(duplicate_words)
purpose_data.write_jsonl(
swe,
[
{"prompt": "Diagnose a unique worker crash"},
{"prompt": first_duplicate},
{"prompt": second_duplicate},
],
)
stage = script_dir / ".artifacts" / "sol-high-reset"
public_dataset = script_dir / ".artifacts" / "dataset-public"
combined_dataset = script_dir / ".artifacts" / "dataset-v1"
@@ -222,7 +233,10 @@ class RebuildSolHighTests(unittest.TestCase):
(item for item in records if item["prompt"] == prompt),
None,
)
record = original or source_record(prompt)
record = original or source_record(
prompt,
"writing" if prompt == second_duplicate else "backendImpl",
)
states.append(
{
"schemaVersion": 1,
@@ -240,10 +254,12 @@ class RebuildSolHighTests(unittest.TestCase):
result = rebuild_sol_high.promote(
stage,
rebuild_sol_high.CONFIRMATION,
excluded_real_lines=[3],
)
self.assertEqual(80, result["publicLabeled"])
self.assertEqual(2, result["realLabeled"])
self.assertEqual(3, result["realLabeled"])
self.assertEqual(3, result["realReviewExclusions"][0]["sourceLine"])
self.assertTrue((combined_dataset / "train.jsonl").is_file())
self.assertGreater(
len(purpose_data.load_jsonl(combined_dataset / "train.jsonl")),
@@ -256,11 +272,42 @@ class RebuildSolHighTests(unittest.TestCase):
"purpose-dataset-sol-high-v2",
promoted_manifest["datasetVersion"],
)
combined_manifest = json.loads(
(combined_dataset / "manifest.json").read_text(encoding="utf-8")
)
self.assertEqual(
result["realReviewExclusions"],
combined_manifest["reviewedRealExclusions"],
)
def test_promotion_requires_exact_confirmation(self):
with self.assertRaisesRegex(purpose_data.DataError, "promotion requires"):
rebuild_sol_high.promote(Path("unused"), None)
def test_reviewed_real_line_exclusions_are_auditable(self):
records = [
source_record("Plan the cache migration", "planning"),
source_record("Review the cache migration", "review"),
]
retained, review = rebuild_sol_high._exclude_reviewed_real_lines(records, [1])
self.assertEqual([records[1]], retained)
self.assertEqual(1, review[0]["sourceLine"])
self.assertEqual("planning", review[0]["purpose"])
self.assertEqual(
"reviewed-near-duplicate-label-conflict", review[0]["reason"]
)
self.assertEqual(64, len(review[0]["promptHash"]))
def test_reviewed_real_line_exclusions_reject_invalid_lines(self):
records = [source_record("Plan the cache migration", "planning")]
with self.assertRaisesRegex(purpose_data.DataError, "duplicate line"):
rebuild_sol_high._exclude_reviewed_real_lines(records, [1, 1])
with self.assertRaisesRegex(purpose_data.DataError, "outside"):
rebuild_sol_high._exclude_reviewed_real_lines(records, [2])
if __name__ == "__main__":
unittest.main()