LabLens / lablens /evaluate.py
Imene-ee's picture
Upload 19 files
dab7db2 verified
Raw History Blame Contribute Delete
4.83 kB
"""Benchmark the extraction + flagging pipeline against synthetic ground truth.
Metrics (per input type: PDF text layer vs degraded photo, and per locale):
recall share of true test rows that were found
precision share of extracted rows that are real rows of the report
value_acc exact value match among found rows
unit_acc unit understood as the same unit (same conversion factor)
range_acc reference range read exactly
flag_acc end-to-end flag equals the ground-truth flag (missing rows count as wrong)
critical_recall critical rows that came out as critical (the safety metric)
dangerous_misses abnormal rows reported as NORMAL (worse than missing: false reassurance)
"""
from __future__ import annotations
import argparse
import json
from collections import defaultdict
from pathlib import Path
from .catalog import load_catalog
from .extract import extract
from .flagger import flag_all
def _same_unit(key: str, a: str | None, b: str | None) -> bool:
td = load_catalog()[key]
fa, fb = td.to_canonical(1.0, a), td.to_canonical(1.0, b)
return fa is not None and fb is not None and abs(fa - fb) < 1e-9
def score_report(truth: dict, path: Path) -> dict:
results, _, _, _ = extract(path)
analysis = flag_all(results)
got = {r.test_key: r for r in analysis.results}
true_keys = {row["test_key"] for row in truth["rows"]}
c = defaultdict(int)
c["rows"] = len(truth["rows"])
c["extracted"] = len(got)
c["true_positive_keys"] = len(true_keys & set(got))
for row in truth["rows"]:
k, exp = row["test_key"], row["expected_flag"]
crit = exp.startswith("critical")
c["critical_rows"] += crit
c["abnormal_rows"] += exp != "normal"
g = got.get(k)
if g is None:
continue
c["found"] += 1
c["value_ok"] += abs(g.value - row["value"]) < 1e-9
c["unit_ok"] += _same_unit(k, g.unit, row["unit"])
c["range_ok"] += (g.ref_low, g.ref_high) == (row["ref_low"], row["ref_high"]) or (
g.ref_source == "catalog" and row["ref_low"] is None and row["ref_high"] is None)
c["flag_ok"] += g.flag.value == exp
c["critical_caught"] += crit and g.flag.value.startswith("critical")
c["dangerous_miss"] += exp != "normal" and g.flag.value == "normal"
return dict(c)
def summarize(counts: list[dict]) -> dict:
t = defaultdict(int)
for c in counts:
for k, v in c.items():
t[k] += v
def ratio(a, b):
return round(t[a] / t[b], 4) if t[b] else None
return {
"reports": len(counts), "rows": t["rows"],
"recall": ratio("found", "rows"),
"precision": ratio("true_positive_keys", "extracted"),
"value_acc": ratio("value_ok", "found"),
"unit_acc": ratio("unit_ok", "found"),
"range_acc": ratio("range_ok", "found"),
"flag_acc": ratio("flag_ok", "rows"),
"critical_recall": ratio("critical_caught", "critical_rows"),
"critical_rows": t["critical_rows"],
"dangerous_misses": t["dangerous_miss"],
"abnormal_rows": t["abnormal_rows"],
}
def run(bench_dir: Path) -> dict:
groups: dict[str, list[dict]] = defaultdict(list)
for jf in sorted(bench_dir.glob("rpt_*.json")):
truth = json.loads(jf.read_text(encoding="utf-8"))
for kind, field in (("pdf", "pdf"), ("photo", "image")):
if field not in truth:
continue
c = score_report(truth, bench_dir / truth[field])
groups[f"{kind}/all"].append(c)
groups[f"{kind}/{truth['locale']}"].append(c)
return {g: summarize(v) for g, v in sorted(groups.items())}
def to_markdown(summary: dict) -> str:
cols = ["reports", "rows", "recall", "precision", "value_acc", "unit_acc", "range_acc",
"flag_acc", "critical_recall", "dangerous_misses"]
out = ["| split | " + " | ".join(cols) + " |", "|" + "---|" * (len(cols) + 1)]
for g, s in summary.items():
cells = []
for k in cols:
v = s[k]
if isinstance(v, float) and k not in ("reports", "rows", "dangerous_misses"):
cells.append(f"{v:.1%}")
elif k == "critical_recall" and v is None:
cells.append("n/a")
else:
cells.append(str(v))
out.append(f"| {g} | " + " | ".join(cells) + " |")
return "\n".join(out)
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("--bench", default="benchmark/reports")
ap.add_argument("--out", default="benchmark/results.json")
a = ap.parse_args()
summary = run(Path(a.bench))
Path(a.out).write_text(json.dumps(summary, indent=2), encoding="utf-8")
print(to_markdown(summary))