File size: 4,686 Bytes
94da461
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""SANDBOX C3 run: judge the 51 temp merged injections with the SAME procedure

as the baseline run (single orientation, samples=3 majority, workers=4, model

from datasetreview/config.yaml). Compares the injected caught-rate against the

baseline overall rate and the 51 host rows' own baseline caught. Writes results

only into the sandbox (out/c3_injected_results.jsonl); canonical files untouched.



Run from repo root:  python -u temp/injection_sandbox/run_c3.py

"""
from __future__ import annotations

import json
import logging
import random
import sys
from collections import Counter
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path

import yaml

for _n in ("azure", "azure.identity", "azure.core.pipeline.policies.http_logging_policy"):
    logging.getLogger(_n).setLevel(logging.WARNING)

ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(ROOT))

from datasetreview import pipelines as P          # noqa: E402
from datasetreview import judge_prompts as J       # noqa: E402
from datasetreview.llm_client import make_judge     # noqa: E402

SAND = Path(__file__).resolve().parent
OUT = SAND / "out"
MERGED = OUT / "_staging_merged_injections.jsonl"
PROV = json.load(open(OUT / "_merged_injections.prov.json", encoding="utf-8"))
BASELINE = ROOT / "datasetreview" / "results" / "new" / "C3.jsonl"

cfg = yaml.safe_load(open(ROOT / "datasetreview" / "config.yaml", encoding="utf-8"))
SAMPLES = 3          # baseline C3 used samples=3 majority
WORKERS = cfg["run"].get("workers", 4)

reals = P.real_trajectories()
pairer = P.make_pairer(reals)
judge = make_judge(cfg["model"])
fakes = [json.loads(l) for l in open(MERGED, encoding="utf-8") if l.strip()]


def judge_one(fake):
    eid = fake["example_id"]
    real = pairer(fake)
    swap = random.Random(eid).random() < 0.5
    msgs = J.build_c3(fake, real, swap=swap)
    key = msgs["answer_key"]
    guesses, err = [], None
    for _ in range(SAMPLES):
        try:
            r = judge.judge({"system": msgs["system"], "user": msgs["user"]})
            guesses.append(r.get("guess"))
        except Exception as e:                 # noqa: BLE001
            err = str(e)
    if not guesses:
        return {"example_id": eid, "error": err, "answer_key": key, "real_id": real.get("example_id")}
    majority = Counter(guesses).most_common(1)[0][0]
    agree = guesses.count(majority) / len(guesses)
    return {"example_id": eid, "answer_key": key, "real_id": real.get("example_id"),
            "guess": majority, "agreement": agree, "n_samples": len(guesses),
            "sample_guesses": guesses, "caught": majority == key, "error": None}


def main():
    print(f"judging {len(fakes)} injected rows  (samples={SAMPLES}, workers={WORKERS}, "
          f"model={cfg['model']['label']})")
    with ThreadPoolExecutor(max_workers=WORKERS) as ex:
        results = list(ex.map(judge_one, fakes))
    results.sort(key=lambda r: r["example_id"])
    (OUT / "c3_injected_results.jsonl").write_text(
        "\n".join(json.dumps(r) for r in results) + "\n", encoding="utf-8")

    ok = [r for r in results if r.get("error") is None and "caught" in r]
    errs = [r for r in results if r.get("error") is not None or "caught" not in r]
    caught = sum(1 for r in ok if r["caught"])
    n = len(ok)

    # baseline comparison
    base = {}
    for line in open(BASELINE, encoding="utf-8"):
        if line.strip():
            d = json.loads(line)
            base[d.get("item_id")] = d.get("caught")
    base_all = [v for v in base.values() if v is not None]
    base_rate = sum(base_all) / len(base_all) if base_all else 0.0
    host_of = {p["example_id"]: p["host_of"] for p in PROV}
    host_caught = [base.get(host_of[r["example_id"]]) for r in ok]
    host_caught = [v for v in host_caught if v is not None]
    host_rate = sum(host_caught) / len(host_caught) if host_caught else 0.0

    print("\n=== C3 RESULT (injected rows) ===")
    print(f"  injected caught: {caught}/{n} = {caught/n:.1%}" if n else "  no valid results")
    print(f"  errors: {len(errs)}")
    print(f"\n  baseline overall (795 rows): {sum(base_all)}/{len(base_all)} = {base_rate:.1%}")
    print(f"  the 51 hosts' own baseline caught: {sum(host_caught)}/{len(host_caught)} = {host_rate:.1%}")
    delta = caught / n - base_rate if n else 0.0
    print(f"\n  injected vs baseline-overall delta: {delta:+.1%}  "
          f"(negative = injected fooled the judge MORE than baseline)")
    print(f"  wrote out/c3_injected_results.jsonl")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())