feat(#000057): bench/score_with_code_judge.py — code-judge rescore tool
Reads an existing sweep JSONL (Opus verdicts already recorded), re-fetches
gold per record via _gold(), runs the code judge on (question_asked,
answer, gold), and emits:
- markdown scorecard: agreement matrix (code × original judge), per-arm
/ per-model / per-variant code-judge tallies, and a residue table of
the JUDGE_ERROR records (the natural input to a later LLM-batch
needle-haystack pass — Opus or Grok);
- JSONL with one row per sweep record (code_verdict + code_rationale
+ code_decision), joinable on (i, arm, model, variant) to the
source sweep.
Zero LLM calls. Reads sweep JSONL + shards read-only. Pairs with the
new --judge switch (a2e9b49): the switch decides what NEW data uses;
this script decides what the ALREADY-COLLECTED data looks like under
the deterministic judge.
Usage (parameter default matches control_sweep.py default fixture):
python -m bench.score_with_code_judge --in <sweep>.jsonl
Currently running against control_sweep_2026-05-19T17-01-17Z.jsonl
(the 2289-record sweep that ran on the prior huge-N pass before the
Opus quota burned out). Output will land at
bench/qa_results/control_sweep_2026-05-19T17-01-17Z_code_judge.{md,jsonl}.
This commit is contained in:
parent
a2e9b49c0a
commit
7913012001
1 changed files with 262 additions and 0 deletions
262
bench/score_with_code_judge.py
Normal file
262
bench/score_with_code_judge.py
Normal file
|
|
@ -0,0 +1,262 @@
|
|||
#!/usr/bin/env python3
|
||||
"""Re-grade an existing #000057 control-sweep JSONL with the code judge.
|
||||
|
||||
fox 2026-05-19: "we burned all our Opus … let's get data first and we
|
||||
can use tools to judge before passing samples to Opus later." The sweep
|
||||
JSONLs from the huge-N run carry every (question, answer) but NOT the
|
||||
gold source (gold is re-fetched from shards at grade time so the JSONL
|
||||
stays compact). This script:
|
||||
|
||||
1. loads a sweep JSONL (the recorded Opus verdicts on each record);
|
||||
2. for each record, re-fetches gold via ``bench.control_ab._gold``
|
||||
using the matching fixture entry (matched by item index ``i``);
|
||||
3. runs ``bench.judge_code.judge`` on (question_asked, answer, gold);
|
||||
4. emits an agreement matrix: code-judge label × opus-judge label,
|
||||
plus a per-arm / per-model / per-variant scorecard;
|
||||
5. emits the JUDGE_ERROR residue list — the records the code judge
|
||||
could not classify deterministically. That residue is the eventual
|
||||
LLM-batch needle-haystack ask. If the residue is small, you do not
|
||||
need a paid second pass; if it's big, you know exactly the size of
|
||||
the spend you would be authorising.
|
||||
|
||||
NO LLM calls, no shard writes. Reads the existing JSONL & shards
|
||||
read-only; writes a markdown scorecard + a JSONL of code-judge verdicts
|
||||
(one row per existing-sweep record, joinable on ``i,arm,model,variant``).
|
||||
|
||||
Usage::
|
||||
|
||||
python -m bench.score_with_code_judge \\
|
||||
--in bench/qa_results/control_sweep_<TS>.jsonl \\
|
||||
--fixture bench/qa_questions_stale_map.json \\
|
||||
--shards-dir ~/.arborist/shards \\
|
||||
--out-md bench/results/code_judge_rescore_<TS>.md \\
|
||||
--out-jsonl bench/qa_results/code_judge_rescore_<TS>.jsonl
|
||||
|
||||
The default fixture matches the sweep's default; if you ran the sweep
|
||||
with a different ``--fixture`` you must pass the same one here.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import sys
|
||||
import time
|
||||
from collections import Counter
|
||||
from pathlib import Path
|
||||
|
||||
ROOT = Path(__file__).resolve().parents[1]
|
||||
sys.path.insert(0, str(ROOT))
|
||||
from bench.control_ab import _gold # noqa: E402
|
||||
from bench.judge_code import judge as code_judge # noqa: E402
|
||||
from bench.judge_code import JUDGE_MODEL as CODE_JUDGE_MODEL # noqa: E402
|
||||
|
||||
|
||||
def _load_jsonl(p: Path) -> list[dict]:
|
||||
"""Tolerant load — same convention as control_sweep._load_recs."""
|
||||
out = []
|
||||
for ln in p.read_text().splitlines():
|
||||
if not ln.strip():
|
||||
continue
|
||||
try:
|
||||
out.append(json.loads(ln))
|
||||
except json.JSONDecodeError:
|
||||
continue
|
||||
return out
|
||||
|
||||
|
||||
def _agree_md(matrix: Counter, code_total: Counter, opus_total: Counter) -> list[str]:
|
||||
labels = ["CORRECT_GROUNDED", "WRONG", "FABRICATED", "ABSTAINED",
|
||||
"JUDGE_ERROR"]
|
||||
L = ["## Agreement matrix (rows = code judge, columns = original judge in sweep)",
|
||||
"",
|
||||
"| code \\ orig | " + " | ".join(labels) + " | row total |",
|
||||
"|" + "|".join(["---"] * (len(labels) + 2)) + "|"]
|
||||
for cl in labels:
|
||||
row = [str(matrix.get((cl, ol), 0)) for ol in labels]
|
||||
L.append(f"| **{cl}** | " + " | ".join(row) + f" | {code_total[cl]} |")
|
||||
L.append("| **col total** | " + " | ".join(
|
||||
str(opus_total[ol]) for ol in labels) + " | |")
|
||||
# Diagonal = exact agreement; off-diagonal = disagreement.
|
||||
diag = sum(matrix.get((lab, lab), 0) for lab in labels)
|
||||
total = sum(matrix.values())
|
||||
if total:
|
||||
L += ["",
|
||||
f"Exact-label agreement: **{diag}/{total} = "
|
||||
f"{diag/total:.1%}**."]
|
||||
return L
|
||||
|
||||
|
||||
def _bucket_str(c: Counter) -> str:
|
||||
tot = sum(c.values()) or 1
|
||||
return (f"CG={c['CORRECT_GROUNDED']} W={c['WRONG']} "
|
||||
f"F={c['FABRICATED']} A={c['ABSTAINED']} "
|
||||
f"JE={c['JUDGE_ERROR']} (n={tot})")
|
||||
|
||||
|
||||
def _scorecard_md(by_key: dict, title: str) -> list[str]:
|
||||
L = [f"## {title}",
|
||||
"",
|
||||
"Each row is the same set of records as the matching column "
|
||||
"of the original sweep — only the judge changed.",
|
||||
"",
|
||||
"| key | code-judge verdicts |",
|
||||
"|---|---|"]
|
||||
for k, c in sorted(by_key.items()):
|
||||
L.append(f"| `{k}` | {_bucket_str(c)} |")
|
||||
return L
|
||||
|
||||
|
||||
def _residue_md(residue: list[dict], cap: int = 30) -> list[str]:
|
||||
L = [f"## JUDGE_ERROR residue (top {min(cap, len(residue))} of "
|
||||
f"{len(residue)})",
|
||||
"",
|
||||
"These are the records the code judge could NOT classify "
|
||||
"deterministically. They are the natural input to a later "
|
||||
"LLM-batch needle-haystack pass (Opus or Grok). Each row "
|
||||
"shows the code-judge rationale + a short answer excerpt so "
|
||||
"the residue is greppable / sortable before any LLM spend.",
|
||||
"",
|
||||
"| i | arm | model | variant | original | code rationale | answer |",
|
||||
"|---|---|---|---|---|---|---|"]
|
||||
for r in residue[:cap]:
|
||||
ans = (r.get("answer") or "").replace("|", "/")[:80]
|
||||
L.append(f"| {r.get('i')} | {r.get('arm')} | {r.get('model')} | "
|
||||
f"{r.get('variant')} | {r.get('verdict')} | "
|
||||
f"{(r.get('code_rationale') or '')[:90]} | {ans} |")
|
||||
return L
|
||||
|
||||
|
||||
def main() -> int:
|
||||
ap = argparse.ArgumentParser()
|
||||
ap.add_argument("--in", dest="in_path", required=True,
|
||||
help="sweep JSONL to re-grade")
|
||||
ap.add_argument("--fixture",
|
||||
default="bench/qa_questions_stale_map.json",
|
||||
help="must match the fixture the sweep was run on "
|
||||
"(records carry item index i, not gold)")
|
||||
ap.add_argument("--shards-dir",
|
||||
default=str(Path.home() / ".arborist" / "shards"))
|
||||
ap.add_argument("--out-md", default="",
|
||||
help="markdown scorecard output path "
|
||||
"(default: alongside --in with _code_judge.md)")
|
||||
ap.add_argument("--out-jsonl", default="",
|
||||
help="JSONL of per-record code-judge verdicts "
|
||||
"(default: alongside --in with _code_judge.jsonl)")
|
||||
a = ap.parse_args()
|
||||
|
||||
in_path = Path(a.in_path)
|
||||
if not in_path.exists():
|
||||
print(f"ABORT: {in_path} not found", file=sys.stderr)
|
||||
return 1
|
||||
out_md = Path(a.out_md) if a.out_md else \
|
||||
in_path.with_name(in_path.stem + "_code_judge.md")
|
||||
out_jsonl = Path(a.out_jsonl) if a.out_jsonl else \
|
||||
in_path.with_name(in_path.stem + "_code_judge.jsonl")
|
||||
|
||||
items = json.loads(Path(a.fixture).read_text())
|
||||
# Index `i` in the sweep records is 1-based (enumerate start=1) and
|
||||
# references items in fixture order; rebuild the lookup table.
|
||||
by_i = {i + 1: it for i, it in enumerate(items)}
|
||||
|
||||
recs = _load_jsonl(in_path)
|
||||
if not recs:
|
||||
print(f"ABORT: {in_path} has no usable records", file=sys.stderr)
|
||||
return 1
|
||||
|
||||
shards_dir = Path(a.shards_dir)
|
||||
gold_cache: dict[int, str] = {}
|
||||
matrix: Counter = Counter() # (code_label, opus_label)
|
||||
code_total: Counter = Counter()
|
||||
opus_total: Counter = Counter()
|
||||
by_arm: dict[str, Counter] = {}
|
||||
by_model: dict[str, Counter] = {}
|
||||
by_variant: dict[str, Counter] = {}
|
||||
residue: list[dict] = []
|
||||
|
||||
t0 = time.time()
|
||||
n_skip_no_gold = 0
|
||||
n_skip_no_fixture_match = 0
|
||||
with open(out_jsonl, "w") as out_log:
|
||||
for r in recs:
|
||||
if r.get("arm") == "skip":
|
||||
continue
|
||||
i = r.get("i")
|
||||
fxr = by_i.get(i)
|
||||
if fxr is None:
|
||||
n_skip_no_fixture_match += 1
|
||||
continue
|
||||
if i not in gold_cache:
|
||||
gold_cache[i] = _gold(shards_dir, fxr.get("shard", ""),
|
||||
fxr["target_root"]) or ""
|
||||
gold = gold_cache[i]
|
||||
if not gold:
|
||||
n_skip_no_gold += 1
|
||||
continue
|
||||
q_asked = r.get("question_asked") or r.get("question_orig") or ""
|
||||
answer = r.get("answer") or ""
|
||||
v = code_judge(q_asked, answer, gold)
|
||||
opus_label = r.get("verdict", "?")
|
||||
matrix[(v.label, opus_label)] += 1
|
||||
code_total[v.label] += 1
|
||||
opus_total[opus_label] += 1
|
||||
by_arm.setdefault(r.get("arm", "?"), Counter())[v.label] += 1
|
||||
by_model.setdefault(r.get("model", "?"), Counter())[v.label] += 1
|
||||
by_variant.setdefault(r.get("variant", "?"), Counter())[v.label] += 1
|
||||
row = {**r,
|
||||
"code_verdict": v.label,
|
||||
"code_rationale": v.rationale,
|
||||
"code_model": v.model,
|
||||
"code_prompt_id": v.prompt_id,
|
||||
"code_decision": v.decision}
|
||||
out_log.write(json.dumps(row, default=str) + "\n")
|
||||
out_log.flush()
|
||||
if v.label == "JUDGE_ERROR":
|
||||
residue.append({**r,
|
||||
"code_rationale": v.rationale})
|
||||
|
||||
dt = time.time() - t0
|
||||
n = sum(matrix.values())
|
||||
print(f" re-graded {n} records in {dt:.1f}s "
|
||||
f"({n/dt:.0f} rec/s) — "
|
||||
f"skip(no_gold)={n_skip_no_gold} "
|
||||
f"skip(no_fixture_match)={n_skip_no_fixture_match}")
|
||||
print(f" per-record JSONL: {out_jsonl}")
|
||||
|
||||
L: list[str] = [
|
||||
f"# Code-judge re-scorecard for `{in_path.name}`",
|
||||
"",
|
||||
f"Generated {time.strftime('%Y-%m-%d %H:%M UTC', time.gmtime())}.",
|
||||
"",
|
||||
f"- Input sweep JSONL: `{in_path}` ({n} records re-graded; "
|
||||
f"skipped {n_skip_no_gold} no-gold + "
|
||||
f"{n_skip_no_fixture_match} fixture-miss).",
|
||||
f"- Fixture: `{a.fixture}` (must match the sweep's --fixture).",
|
||||
f"- Shards: `{shards_dir}`",
|
||||
f"- Code judge: `{CODE_JUDGE_MODEL}` "
|
||||
"(see `bench/judge_code.py`).",
|
||||
"",
|
||||
"**No LLM calls were made by this rescore.** The code-judge "
|
||||
"labels are deterministic; the original (Opus) labels are the "
|
||||
"ones already on disk in the sweep JSONL — they are not "
|
||||
"re-run here. JUDGE_ERROR residue is the eventual "
|
||||
"LLM-batch needle-haystack input (see fox 2026-05-19).",
|
||||
"",
|
||||
]
|
||||
L += _agree_md(matrix, code_total, opus_total)
|
||||
L += [""]
|
||||
L += _scorecard_md(by_arm, "Code-judge scorecard by arm")
|
||||
L += [""]
|
||||
L += _scorecard_md(by_model, "Code-judge scorecard by model")
|
||||
L += [""]
|
||||
L += _scorecard_md(by_variant, "Code-judge scorecard by variant")
|
||||
L += [""]
|
||||
L += _residue_md(residue)
|
||||
|
||||
out_md.parent.mkdir(parents=True, exist_ok=True)
|
||||
out_md.write_text("\n".join(L) + "\n")
|
||||
print(f" markdown scorecard: {out_md}")
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
Loading…
Add table
Add a link
Reference in a new issue