File size: 5,294 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 | """Join every C3 result row to its confuser story and the EXACT text C3 judged.
For each caught row we emit: item_id, tool, real_id, confidence, the judge's
reasoning, the query, a compact history view, the rendered C3 conversation, and
tell tags (keyword-derived from the reasoning). Also emits an uncaught set so we
can study what already fools C3.
Run from repo root: python -u temp/story_remediation/build_join.py
"""
from __future__ import annotations
import json
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(ROOT))
from datasetreview import judge_prompts as J # noqa: E402
OUT = Path(__file__).resolve().parent / "out"
C3 = ROOT / "datasetreview" / "results" / "new" / "C3.jsonl"
TRAJ = ROOT / "distractor_generation_2" / "datasets" / "n100" / "trajectories.jsonl"
def _rows(p: Path):
return [json.loads(l) for l in p.read_text(encoding="utf-8").splitlines() if l.strip()]
TELLS = {
"bundled_checklist": ("bundl", "packs", "checklist", "enumerat", "multiple distinct",
"multi-task", "one turn", "one message", "exercise many",
"exercise multiple", "one-to-one", "stack", "all at once",
"everything at once", "comprehensive", "front-load"),
"no_verification": ("verificat", "authenticat", "identity", "without verif",
"skips ident", "glossed", "no auth"),
"over_precise_ids": ("precise", "exact order", "exact item", "id-laden", "unprompted",
"conveniently", "verbatim", "recite", "specific order number"),
"agent_burst_no_confirm": ("fires", "burst", "no confirmation", "no intermediate",
"without confirm", "batch of tool", "silently", "no clarif",
"many tool call", "rapid sequence"),
"terse_closer": ("taken care of", "all set", "that's handled", "done.", "what else",
"formulaic", "terse"),
"duplicate_request": ("redundant", "restate", "re-asks", "re-state", "twice",
"double-statement", "identical repeated", "repeats the"),
"invented_capability": ("unusual", "atypical", "don't typically", "invent",
"doesn't typically", "not typically", "uncommon", "rarely"),
"sequential_orders": ("sequential", "conveniently formatted", "clean order number"),
"scripted_flavor": ("scripted", "staged", "stagey", "stilted", "canned", "test prompt",
"task prompt", "flavor text", "persona"),
}
def tag(reason: str) -> list[str]:
r = reason.lower()
return [k for k, ws in TELLS.items() if any(w in r for w in ws)]
def hist_view(story: dict):
out = []
for m in story.get("history") or []:
if m.get("role") == "tool":
continue
out.append({"role": m.get("role"),
"content": (m.get("content") or "").strip(),
"n_tool_calls": len(m.get("tool_calls") or [])})
return out
def main() -> int:
c3 = _rows(C3)
traj = {t["example_id"]: t for t in _rows(TRAJ)}
caught, fooled = [], []
missing = 0
for r in c3:
iid = r["item_id"]
story = traj.get(iid)
if story is None:
missing += 1
continue
parts = iid.split("-")
tool = parts[1] if len(parts) >= 3 else "?"
rec = {
"item_id": iid,
"tool": tool,
"real_id": r.get("real_id"),
"caught": bool(r.get("caught")),
"confidence": (r.get("result") or {}).get("confidence"),
"reasoning": (r.get("result") or {}).get("reasoning") or "",
"tells": tag((r.get("result") or {}).get("reasoning") or ""),
"n_calls": len(story.get("calls") or []),
"n_hist_turns": len(hist_view(story)),
"query": (story.get("query") or story.get("retrieval_text") or "").strip(),
"history": hist_view(story),
"c3_view": J.render_trajectory(story, blind_tools=True,
include_metadata=False, conversation_only=True),
}
(caught if rec["caught"] else fooled).append(rec)
(OUT / "caught_rows.jsonl").write_text(
"\n".join(json.dumps(r, ensure_ascii=False) for r in caught) + "\n", encoding="utf-8")
(OUT / "fooled_rows.jsonl").write_text(
"\n".join(json.dumps(r, ensure_ascii=False) for r in fooled) + "\n", encoding="utf-8")
from collections import Counter
tally = Counter(t for r in caught for t in r["tells"])
notag = sum(1 for r in caught if not r["tells"])
print(f"caught={len(caught)} fooled={len(fooled)} missing_story={missing}")
print(f"caught rows with NO tell tag: {notag}")
print("tell frequency among caught:")
for k, v in tally.most_common():
print(f" {v:4d} {k}")
# history presence
openers = sum(1 for r in caught if r["n_hist_turns"] == 0)
print(f"caught openers (no history): {openers} | caught with history: {len(caught)-openers}")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|