File size: 10,399 Bytes
2e511b5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 | """Compare the two Legora runs (legora-1 vs legora-2) in structure and performance.
The 2026-08-01 Legora export carried the full question set twice; the ingest
stores the left group as ``legora-1`` and the right group as ``legora-2``.
This script contrasts the two runs — coverage, per-field fill rates, inter-run
agreement, accuracy against the Goldensets, and example disagreements — and
writes a markdown report.
uv run python scripts/compare_legora_runs.py \
[--out data/analysis/legora_run_comparison.md]
Run it after `legex-refusals-apply` so the report reflects the cleaned data.
"""
import argparse
import sys
from collections import Counter, defaultdict
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from legex import published # noqa: E402
from legex.config import settings # noqa: E402
from legex.evaluation.comparison import ( # noqa: E402
derived,
normalise,
values_agree,
)
from legex.evaluation.scoring import _read_goldenset_rows, score_country # noqa: E402
from legex.legora import LEGORA_FIELDS # noqa: E402
from legex.utils import inference_path, read_inference_jsonl # noqa: E402
MODEL_A, MODEL_B = "legora-1", "legora-2"
FIELDS = tuple(f for f in LEGORA_FIELDS if f != "case_id")
DEFAULT_OUT = Path("data/analysis/legora_run_comparison.md")
MAX_EXAMPLES_PER_FIELD = 3
def _inference_file(cc: str, prompt_version: str, source: str, model: str,
inference_dir: Path | None) -> Path:
if inference_dir is not None:
return published.inference_file(inference_dir, cc, model)
return inference_path(cc, prompt_version, source, model)
def _ccs(prompt_version: str, source: str, inference_dir: Path | None) -> list[str]:
root = inference_dir if inference_dir is not None else settings.data_dir
out = []
for d in sorted(Path(root).iterdir()):
if not d.is_dir():
continue
cc = d.name
if all(_inference_file(cc, prompt_version, source, m, inference_dir).exists()
for m in (MODEL_A, MODEL_B)):
out.append(cc)
return out
def _by_case(cc: str, prompt_version: str, source: str, model: str,
inference_dir: Path | None) -> dict[str, dict]:
rows = read_inference_jsonl(_inference_file(cc, prompt_version, source, model, inference_dir))
return {str(r.get("case_id")): r for r in rows if r.get("case_id")}
def _pct(num: int, den: int) -> str:
return f"{100 * num / den:.1f}%" if den else "–"
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
parser.add_argument("--prompt_version", default="v3")
parser.add_argument("--source", choices=("full_text", "pdf"), default="full_text")
parser.add_argument("--out", type=Path, default=DEFAULT_OUT)
parser.add_argument("--gold-dir", type=Path, default=None,
help="published goldenset data directory (XLSX workbooks otherwise)")
parser.add_argument("--inference-dir", type=Path, default=None,
help="published inference data directory (working files otherwise)")
args = parser.parse_args()
ccs = _ccs(args.prompt_version, args.source, args.inference_dir)
if not ccs:
raise SystemExit("no country has both legora-1 and legora-2 files")
# ---- collect ------------------------------------------------------------
coverage_rows: list[tuple[str, int, int, int]] = [] # cc, n_a, n_b, overlap
fill = {m: Counter() for m in (MODEL_A, MODEL_B)} # field -> filled
n_pairs = Counter() # field -> paired rows
identical = Counter() # field -> normalised equal
tolerant = Counter() # field -> values_agree
both_filled = Counter()
both_filled_agree = Counter()
only_a = Counter()
only_b = Counter()
examples: dict[str, list[tuple[str, str, str, str, str]]] = defaultdict(list)
inference_dates = set()
for cc in ccs:
a = _by_case(cc, args.prompt_version, args.source, MODEL_A, args.inference_dir)
b = _by_case(cc, args.prompt_version, args.source, MODEL_B, args.inference_dir)
shared = sorted(set(a) & set(b))
coverage_rows.append((cc, len(a), len(b), len(shared)))
gold_lookup: dict[str, dict[str, str]] = {}
if shared:
if args.gold_dir is not None:
_, gold_lookup = published.load_gold_labels(args.gold_dir, cc)
else:
_, gold_lookup = _read_goldenset_rows(cc)
for cid in shared:
ra, rb = a[cid], b[cid]
inference_dates.update(filter(None, (ra.get("inference_date"), rb.get("inference_date"))))
for field in FIELDS:
va, vb = normalise(ra.get(field)), normalise(rb.get(field))
n_pairs[field] += 1
fill[MODEL_A][field] += bool(va)
fill[MODEL_B][field] += bool(vb)
if va and not vb:
only_a[field] += 1
if vb and not va:
only_b[field] += 1
if va == vb:
identical[field] += 1
if values_agree(va, vb, field):
tolerant[field] += 1
if va and vb:
both_filled[field] += 1
if values_agree(va, vb, field):
both_filled_agree[field] += 1
elif len(examples[field]) < MAX_EXAMPLES_PER_FIELD:
gv = gold_lookup.get(normalise(cid), {}).get(field, "")
examples[field].append((cc, cid, va, vb, gv))
# ---- vs gold ------------------------------------------------------------
counters = {m: {f: Counter() for f in FIELDS} for m in (MODEL_A, MODEL_B)}
per_cc_acc: dict[str, dict[str, tuple[int, int]]] = defaultdict(dict) # cc -> model -> (correct, n)
for cc in ccs:
for m in (MODEL_A, MODEL_B):
scored = score_country(cc, args.prompt_version, args.source, m, verbose=False,
gold_dir=args.gold_dir, inference_dir=args.inference_dir)
if scored is None:
continue
col_counters, _stats = scored
correct = n = 0
for field in FIELDS:
c = col_counters.get(field)
if c is None:
continue
counters[m][field].update(c)
correct += c["tp"] + c["tn"]
n += sum(c.values())
per_cc_acc[cc][m] = (correct, n)
# ---- render -------------------------------------------------------------
lines: list[str] = []
w = lines.append
dates = ", ".join(sorted(inference_dates)) or "unknown"
w("# Legora run comparison: legora-1 vs legora-2")
w("")
w(f"The 2026-08-01 Legora tabular-review export (`data/raw/legora_2026-08-01.xlsx`, "
f"inference_date {dates}) contains the question set twice; `legora-1` is the left "
f"column group, `legora-2` the right one. Generated by `scripts/compare_legora_runs.py`.")
w("")
w("## Coverage per country")
w("")
w("| cc | legora-1 rows | legora-2 rows | shared |")
w("|---|---|---|---|")
for cc, na, nb, sh in coverage_rows:
w(f"| {cc} | {na} | {nb} | {sh} |")
total_a = sum(r[1] for r in coverage_rows)
total_b = sum(r[2] for r in coverage_rows)
total_s = sum(r[3] for r in coverage_rows)
w(f"| **total** | **{total_a}** | **{total_b}** | **{total_s}** |")
w("")
w("## Per-field fill rates and inter-run agreement")
w("")
w("Agreement uses the evaluation's tolerant comparator (`values_agree`); "
"*identical* is exact string equality after normalisation. *only 1/only 2* "
"count cells filled by one run and empty in the other.")
w("")
w("| field | filled 1 | filled 2 | identical | agree (tolerant) | agree when both filled | only 1 | only 2 |")
w("|---|---|---|---|---|---|---|---|")
for f in FIELDS:
n = n_pairs[f]
w(f"| {f} | {_pct(fill[MODEL_A][f], n)} | {_pct(fill[MODEL_B][f], n)} "
f"| {_pct(identical[f], n)} | {_pct(tolerant[f], n)} "
f"| {_pct(both_filled_agree[f], both_filled[f])} "
f"| {only_a[f]} | {only_b[f]} |")
w("")
w("## Accuracy against the Goldensets")
w("")
w("Cell buckets from the standard scoring (`classify_cell`): accuracy = (tp+tn)/n, "
"precision/recall/F1 as in the analysis pipeline.")
w("")
w("| field | acc 1 | acc 2 | Δ acc | F1 1 | F1 2 | recall 1 | recall 2 |")
w("|---|---|---|---|---|---|---|---|")
for f in FIELDS:
accs = {}
stats = {}
for m in (MODEL_A, MODEL_B):
c = counters[m][f]
n = sum(c.values())
accs[m] = (c["tp"] + c["tn"]) / n if n else 0.0
stats[m] = derived(c)
w(f"| {f} | {accs[MODEL_A]:.3f} | {accs[MODEL_B]:.3f} "
f"| {accs[MODEL_B] - accs[MODEL_A]:+.3f} "
f"| {stats[MODEL_A][2]:.3f} | {stats[MODEL_B][2]:.3f} "
f"| {stats[MODEL_A][1]:.3f} | {stats[MODEL_B][1]:.3f} |")
w("")
w("### Per-country accuracy (all fields pooled)")
w("")
w("| cc | acc legora-1 | acc legora-2 | Δ |")
w("|---|---|---|---|")
for cc in ccs:
accs = {}
for m in (MODEL_A, MODEL_B):
correct, n = per_cc_acc.get(cc, {}).get(m, (0, 0))
accs[m] = correct / n if n else 0.0
w(f"| {cc} | {accs[MODEL_A]:.3f} | {accs[MODEL_B]:.3f} | {accs[MODEL_B] - accs[MODEL_A]:+.3f} |")
w("")
w("## Example disagreements (both runs filled, values differ)")
w("")
w("| field | cc | case_id | legora-1 | legora-2 | gold |")
w("|---|---|---|---|---|---|")
def _md(s: str) -> str:
return s.replace("|", "\\|").replace("\n", " ")[:80]
for f in FIELDS:
for cc, cid, va, vb, gv in examples[f]:
w(f"| {f} | {cc} | {_md(cid)} | {_md(va)} | {_md(vb)} | {_md(gv)} |")
w("")
args.out.parent.mkdir(parents=True, exist_ok=True)
args.out.write_text("\n".join(lines), encoding="utf-8")
print(f"wrote {args.out} ({len(ccs)} countries, {total_s} shared rows)")
if __name__ == "__main__":
main()
|