code / scripts /alt_test_decomposition.py
anonymous
[code] Reproduction bundle.
2e511b5
Raw
History Blame Contribute Delete
10.8 kB
"""Win / tie / loss decomposition of the Alternative Annotator Test (AAT).
The AAT headline (``alt_test_pooled.csv``, ANALYSIS.md §3.2) reports the winning
rate ω and the advantage probability ρ. Inside the test, an instance is credited
to the candidate whenever it scores *at least as well* as the held-out human
(the indicator is ``1[s_llm >= s_human]``), so ties count for the candidate.
That is the correct convention for the question the AAT asks — *can this model
stand in for a human annotator?* — but it makes ρ unusable for the different
question *is this model better than a human?*. This script keeps the AAT's
leave-one-out comparison and, instead of collapsing it into ρ, writes out the
raw counts ρ is built from:
for each held-out human j and each instance i = (judgment, variable)
s_llm = mean tolerant agreement of the candidate with the 2 remaining humans
s_human = mean tolerant agreement of human j with the same 2 humans
-> llm_better (s_llm > s_human)
tie (s_llm = s_human)
human_better(s_llm < s_human)
rho_alttest = (llm_better + tie) / n <- the AAT's own definition
rho_tiebroken = (llm_better + tie / 2) / n <- ties split evenly
A tie only means "same score against the same two references", so it is worth
being explicit about what it contains. Ties are split two independent ways:
by score level (with two references a score is 0, ½ or 1)
tie_at_1 both matched both references — everybody agrees
tie_at_half each matched exactly one reference; this can only happen
when the two reference experts contradict each other, which
caps every possible score at ½
tie_at_0 neither matched either reference — equally wrong, and still
credited to the candidate by ρ
by whether the candidate actually gave the held-out human's answer
tie_same candidate ≈ held-out human (they really do agree)
tie_diff candidate ≉ held-out human — they gave *different* answers
that happen to be equally close to the references, so ρ
records a candidate win on a genuine disagreement
``refs_disagree`` counts, for context, the comparisons whose two reference
experts do not agree with each other in the first place.
This does not need the upstream AltTest clone: it is a direct, auditable
re-implementation of the comparison the reference implementation performs, and
it reproduces the reference ρ to within 0.02 on every jurisdiction.
Usage
-----
uv run python scripts/alt_test_decomposition.py \
[--countries ge,sg,tw] [--out data/analysis/iaa]
"""
import argparse
import csv
import sys
from collections import defaultdict
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(REPO_ROOT))
from legex.analysis.countries import CORE_COUNTRIES
from legex.analysis.iaa import (
FREE_TEXT_FIELDS,
load_candidate_annotations,
load_human_annotations,
)
from legex.evaluation import values_agree
# The five released candidates the paper's AAT covers. The AAT was frozen
# before the harvey-2 ingest; adding it would change the shipped CSVs and
# ANALYSIS.md, so it stays out deliberately.
MODELS = ("gpt-5.4-mini", "gemini/gemini-3.1-flash-lite", "harvey", "legora-1", "legora-2")
COUNTS = (
"refs_disagree", "llm_better", "human_better",
"tie", "tie_same", "tie_diff", "tie_at_1", "tie_at_half", "tie_at_0",
)
COLUMNS = [
"candidate", "country", "variant", "n_comparisons",
*COUNTS,
"rho_alttest", "rho_tiebroken",
]
def score(prediction: str, references: list[str], field: str) -> float:
"""Mean tolerant agreement — the AAT scoring function used in LEGEX."""
return sum(values_agree(prediction, r, field) for r in references) / len(references)
def decompose(
countries: list[str],
model: str,
gold_dir: Path | None = None,
inference_dir: Path | None = None,
) -> list[dict]:
"""One row per (country, variant) with the win/tie/loss counts."""
humans = load_human_annotations(countries, gold_dir=gold_dir)
candidate = load_candidate_annotations(
countries, "v3", "full_text", model, inference_dir=inference_dir
)
fields = sorted(
{f for fmap in humans.values() for f in fmap if f not in FREE_TEXT_FIELDS}
)
# country -> case_id -> annotator -> {field: value}
by_case: dict[str, dict[str, dict[str, dict[str, str]]]] = defaultdict(
lambda: defaultdict(dict)
)
for (annotator, cc, case_id), fmap in humans.items():
by_case[cc][case_id][annotator] = fmap
candidate_labels = {(cc, case_id): fmap for (_, cc, case_id), fmap in candidate.items()}
rows: list[dict] = []
for cc in countries:
annotators = sorted({an for (an, c, _) in humans if c == cc})
if len(annotators) < 3:
print(f"[{cc}] only {len(annotators)} annotators — skipped", file=sys.stderr)
continue
for variant in ("all", "nontrivial"):
counts: dict[str, int] = defaultdict(int)
for case_id, case_annotators in by_case[cc].items():
# Keep only judgments all three experts labelled, so every
# leave-one-out comparison has exactly two reference annotators
# (the reference implementation's min_humans_per_instance=2).
if len(case_annotators) < 3:
continue
names = sorted(case_annotators)
llm_labels = candidate_labels.get((cc, case_id), {})
# No candidate output for this judgment (e.g. Legora's empty
# Georgia export): skip, mirroring the pooled AAT runner.
if not llm_labels:
continue
for field in fields:
human_values = {n: case_annotators[n].get(field, "") for n in names}
llm_value = llm_labels.get(field, "")
for held_out in names:
references = [human_values[n] for n in names if n != held_out]
# Non-trivial: drop comparisons whose reference is empty
# throughout — there is nothing to be right or wrong about.
if variant == "nontrivial" and not any(references):
continue
if not values_agree(references[0], references[1], field):
counts["refs_disagree"] += 1
s_human = score(human_values[held_out], references, field)
s_llm = score(llm_value, references, field)
if s_llm > s_human:
counts["llm_better"] += 1
elif s_llm < s_human:
counts["human_better"] += 1
else:
counts["tie"] += 1
counts[{1.0: "tie_at_1", 0.5: "tie_at_half"}.get(s_llm, "tie_at_0")] += 1
same = values_agree(llm_value, human_values[held_out], field)
counts["tie_same" if same else "tie_diff"] += 1
n = counts["llm_better"] + counts["tie"] + counts["human_better"]
if not n:
continue
rows.append({
"candidate": model,
"country": cc,
"variant": variant,
"n_comparisons": n,
**{c: counts[c] for c in COUNTS},
"rho_alttest": round((counts["llm_better"] + counts["tie"]) / n, 4),
"rho_tiebroken": round((counts["llm_better"] + counts["tie"] / 2) / n, 4),
})
return rows
def main() -> None:
ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
ap.add_argument("--countries", default=",".join(CORE_COUNTRIES),
help="comma-separated country codes (default: the 8 core "
"jurisdictions, all with 3 independent annotators)")
ap.add_argument("--out", type=Path, default=Path("data/analysis/iaa"),
help="output directory for alt_test_decomposition.csv")
ap.add_argument("--gold-dir", type=Path, default=None,
help="read annotations from published goldenset JSONL under "
"this directory instead of the XLSX workbooks")
ap.add_argument("--inference-dir", type=Path, default=None,
help="read candidate predictions from published inference "
"JSONL under this directory instead of the working files")
args = ap.parse_args()
countries = [c.strip() for c in args.countries.split(",") if c.strip()]
rows = [
r for model in MODELS
for r in decompose(countries, model, args.gold_dir, args.inference_dir)
]
if not rows:
raise SystemExit("no comparisons — need 3+ annotators in at least one country")
out_csv = args.out / "alt_test_decomposition.csv"
out_csv.parent.mkdir(parents=True, exist_ok=True)
with out_csv.open("w", encoding="utf-8", newline="") as f:
writer = csv.DictWriter(f, fieldnames=COLUMNS)
writer.writeheader()
writer.writerows(rows)
for model in MODELS:
pooled: dict[str, int] = defaultdict(int)
for r in rows:
if r["candidate"] == model and r["variant"] == "all":
pooled["n_comparisons"] += r["n_comparisons"]
for key in COUNTS:
pooled[key] += r[key]
n = pooled["n_comparisons"]
if not n:
continue
print(
f"{model:<30} n={n:<5} "
f"better {pooled['llm_better']:>4} ({pooled['llm_better'] / n:.0%}) "
f"tie {pooled['tie']:>4} ({pooled['tie'] / n:.0%}) "
f"worse {pooled['human_better']:>4} ({pooled['human_better'] / n:.0%}) "
f"rho={(pooled['llm_better'] + pooled['tie']) / n:.2f} "
f"(ties split {(pooled['llm_better'] + pooled['tie'] / 2) / n:.2f})\n"
f"{'':<30} ties: same answer as expert {pooled['tie_same']} "
f"({pooled['tie_same'] / pooled['tie']:.0%}), different answer "
f"{pooled['tie_diff']} ({pooled['tie_diff'] / pooled['tie']:.0%}); "
f"at 1 {pooled['tie_at_1']}, at ½ {pooled['tie_at_half']}, at 0 {pooled['tie_at_0']}; "
f"references conflict in {pooled['refs_disagree']} of {n} comparisons "
f"({pooled['refs_disagree'] / n:.0%})"
)
print(f"decomposition -> {out_csv}")
if __name__ == "__main__":
main()