nftbench / src /tests /test_question_audit_release.py
OniReimu's picture
NFTBench 1.0.3 (part 2)
b9c9058 verified
Raw History Blame Contribute Delete
1.82 kB
"""The released clarification-question audit reproduces audit.json from its own files."""
from __future__ import annotations
import importlib.util
import json
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parents[2]
spec = importlib.util.spec_from_file_location("question_audit", REPO_ROOT / "scripts" / "question_audit.py")
QA = importlib.util.module_from_spec(spec)
spec.loader.exec_module(QA)
def test_audit_reproduces_reference():
res, reference = QA.recompute(QA.AUDIT_DIR)
assert QA.mismatches(res, reference) == []
assert (res["n_items"], res["agreement"]["three_way"]["agree"]) == (93, 92)
models = res["models"].values()
assert sum(m["identifies_successes"] for m in models) == 83
assert sum(m["clarify_successes"] for m in models) == 87
assert round(res["models"]["gpt-oss-120b"]["S_q"], 3) == 0.407
assert round(res["models"]["Mistral Small 4"]["S_q"], 3) == 0.330
def test_changed_label_is_reported(tmp_path):
for f in QA.AUDIT_DIR.iterdir():
if f.is_file():
(tmp_path / f.name).write_bytes(f.read_bytes())
adj = tmp_path / "adjudication.jsonl"
adj.write_text(adj.read_text(encoding="utf-8").replace('"identifies"', '"other"'), encoding="utf-8")
assert QA.mismatches(*QA.recompute(tmp_path))
def test_changed_reference_field_is_reported(tmp_path):
for f in QA.AUDIT_DIR.iterdir():
if f.is_file():
(tmp_path / f.name).write_bytes(f.read_bytes())
ref = (tmp_path / "audit.json")
doc = json.loads(ref.read_text(encoding="utf-8"))
doc["agreement"]["identifies_vs_not"]["agree"] = 0
ref.write_text(json.dumps(doc), encoding="utf-8")
assert QA.mismatches(*QA.recompute(tmp_path)) == [
"audit.json.agreement.identifies_vs_not.agree: recomputed 92, audit.json 0"
]