File size: 12,138 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 | """Validates and cleans the inference (Gemini, ChatGPT, Harvey) JSONL data.
We try to auto-recover values that do not match our model, else we review it by hand.
Every non-empty prediction cell is resolved to one of:
* valid — a good value already → kept,
* recovered — a value parseable from prose (e.g. `20'000`→`20000`, `1.0\n\n…`→`1.0`)
→ canonicalised automatically (deterministic, score-neutral),
* review — non-empty but neither valid nor recoverable (refusal, prose with no
value, misattribution, bad ISIC) → sent to the review XLSX for a human.
We perform this in two steps to allow for manual review.
1. `legex-refusals-scan` Write every review cell to a review XLSX.
2. `legex-refusals-apply` Applies the decisions to the JSONL.
Traceability fields on every inference record show the original values:
* comment: one sentence per changed field, or null if none: "The <field> was sanitized from '<old>' to '<new>'."
(removals use "empty (removed)" as the new value)
* original_input: JSON object mapping each changed field to its 1:1 original value, the string "{}" when nothing changed.
This is the same format as for the goldenset workbooks.
"""
import argparse
import json
import logging
import re
import sys
from collections import defaultdict
from pathlib import Path
import openpyxl
from legex.config import settings
from legex.evaluation.comparison import classify_cell, is_label_column, normalise, resolve
from legex.utils import goldenset_path, goldenset_sheet, read_inference_jsonl
log = logging.getLogger(__name__)
_INFERENCE_RE = re.compile(r"^Goldenset_.+_v\d+_(?:full_text|pdf)_(.+)\.jsonl$")
DEFAULT_XLSX = Path("data/analysis/quality/inference_data_review.xlsx")
CHANGELOG = Path("data/analysis/quality/inference_cleaning_changelog.jsonl")
_XLSX_HEADER = [
"model", "country", "case_id", "field", "current_value", "reason",
"gold_value", "before_bucket", "after_if_emptied", "corrected_value",
]
def format_comment(changes: dict[str, tuple[str, str]]) -> str | None:
"""Human note for changed fields (field -> (old, new)); None if nothing changed."""
if not changes:
return None
return " ".join(
f"The {field} was sanitized from '{old}' to "
f"'{new if new else 'empty (removed)'}'."
for field, (old, new) in changes.items()
)
def format_original_input(changes: dict[str, tuple[str, str]]) -> str:
"""JSON of {field: original value} for changed fields; '{}' when none."""
return json.dumps({f: old for f, (old, _) in changes.items()}, ensure_ascii=False) if changes else "{}"
def _inference_files(models: set[str] | None = None) -> list[Path]:
"""All inference JSONL files, optionally restricted to the given model models."""
return sorted(
p for p in settings.data_dir.glob("*/Goldenset_*_v*_*.jsonl")
if _INFERENCE_RE.match(p.name)
and (models is None or _model_slug(p) in models)
)
def _model_slug(path: Path) -> str:
return _INFERENCE_RE.match(path.name).group(1)
def _gold_labels(cc: str) -> dict[str, dict[str, str]]:
"""case_id -> {field: normalised gold value}, or {} if no goldenset."""
gs = goldenset_path(cc)
if not gs or not gs.exists():
return {}
ws = goldenset_sheet(openpyxl.load_workbook(gs, read_only=True, data_only=True))
rows = ws.iter_rows(values_only=True)
header = [str(c) if c is not None else "" for c in next(rows)]
ci = header.index("case_id")
out: dict[str, dict[str, str]] = {}
for row in rows:
if row[ci] in (None, ""):
continue
cells = dict(zip(header, row))
out[str(row[ci]).strip()] = {c: normalise(cells.get(c)) for c in header if is_label_column(c)}
return out
def scan(
out: Path = DEFAULT_XLSX,
models: set[str] | None = None,
conflicts: Path | None = None,
) -> int:
"""Write every `review` cell across the inference JSONL to a review XLSX.
"""
gold_cache: dict[str, dict] = {}
rows_out: list[list] = []
seen_cells: set[tuple[str, str, str, str]] = set()
for path in _inference_files(models):
cc = path.parent.name
model = _model_slug(path)
gold = gold_cache.setdefault(cc, _gold_labels(cc))
for rec in read_inference_jsonl(path):
cid = (rec.get("case_id") or "").strip()
for field, value in rec.items():
if not is_label_column(field):
continue
status, _canon, reason = resolve(value, field)
if status != "review":
continue
gv = gold.get(cid, {}).get(field, "")
rows_out.append([
model, cc, cid, field, str(value), reason, gv,
classify_cell(gv, normalise(value), field),
classify_cell(gv, "", field), "",
])
seen_cells.add((model, cc, cid, field))
if conflicts is not None and conflicts.exists():
with open(conflicts, encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
c = json.loads(line)
model, cc, cid, field = c["model"], c["country"], c["case_id"], c["field"]
if models is not None and model not in models:
continue
if (model, cc, cid, field) in seen_cells:
continue # already flagged by resolve(); one row per cell
gold = gold_cache.setdefault(cc, _gold_labels(cc))
gv = gold.get(cid, {}).get(field, "")
kept = c["kept"]
rows_out.append([
model, cc, cid, field, kept,
f"duplicate-run conflict; alternative: '{c['alternative']}'", gv,
classify_cell(gv, normalise(kept), field),
classify_cell(gv, "", field), kept,
])
rows_out.sort(key=lambda r: (r[0], r[5], r[1], r[2]))
wb = openpyxl.Workbook()
ws = wb.active
ws.title = "review"
ws.append(_XLSX_HEADER)
for r in rows_out:
ws.append(r)
out.parent.mkdir(parents=True, exist_ok=True)
wb.save(out)
log.info(f"flagged {len(rows_out)} cell(s) for review -> {out}")
return len(rows_out)
def _load_review(xlsx: Path) -> dict[tuple[str, str], dict[str, dict[str, str]]]:
"""Reviewed XLSX -> {(model, country): {case_id: {field: corrected_value}}}.
corrected_value is "" when the reviewer left it blank (⇒ empty the cell)."""
ws = openpyxl.load_workbook(xlsx, read_only=True).active
rows = ws.iter_rows(values_only=True)
hdr = list(next(rows))
mi, ci, ii, fi, cv = (hdr.index(x) for x in
("model", "country", "case_id", "field", "corrected_value"))
out: dict[tuple[str, str], dict[str, dict[str, str]]] = defaultdict(lambda: defaultdict(dict))
for r in rows:
if r[mi] is None:
continue
out[(str(r[mi]), str(r[ci]))][str(r[ii])][str(r[fi])] = "" if r[cv] is None else str(r[cv])
return out
def apply(xlsx: Path = DEFAULT_XLSX, models: set[str] | None = None) -> None:
"""Auto-canonicalise recovered cells, apply reviewed decisions, stamp provenance.
"""
reviewed = _load_review(xlsx) if xlsx.exists() else {}
log_rows: list[list[str]] = []
n_recovered = n_reviewed = 0
for path in _inference_files(models):
cc, model = path.parent.name, _model_slug(path)
decisions = reviewed.get((model, cc), {})
records = read_inference_jsonl(path)
for rec in records:
cid = (rec.get("case_id") or "").strip()
try:
prev_orig = json.loads(rec.get("original_input") or "{}")
except (ValueError, TypeError):
prev_orig = {}
row_decisions = decisions.get(cid, {})
changes: dict[str, tuple[str, str]] = {}
for field in list(rec):
if not is_label_column(field):
continue
raw = prev_orig[field] if field in prev_orig else rec.get(field) # true original
if raw in (None, ""):
rec[field] = None
continue
if field in row_decisions: # human decision
new = row_decisions[field]
rec[field] = new or None
if str(raw) != (new or ""):
changes[field] = (str(raw), new)
n_reviewed += 1
continue
status, canon, _ = resolve(raw, field) # auto path
if status == "recovered":
rec[field] = canon
if str(raw) != canon:
changes[field] = (str(raw), canon)
n_recovered += 1
else:
rec[field] = raw # valid / undecided review → keep
rec.pop("comment", None)
rec.pop("original_input", None)
rec["comment"] = format_comment(changes)
rec["original_input"] = format_original_input(changes)
for f, (old, new) in changes.items():
log_rows.append({
"model": model, "country": cc, "case_id": cid, "field": f,
"before": old, "after": new,
"kind": "reviewed" if f in row_decisions else "recovered",
})
with open(path, "w", encoding="utf-8") as fh:
for rec in records:
fh.write(json.dumps(rec, ensure_ascii=False) + "\n")
if models is not None and CHANGELOG.exists(): # keep other models' history
with open(CHANGELOG, encoding="utf-8") as f:
log_rows.extend(
row for row in map(json.loads, filter(str.strip, f))
if row.get("model") not in models
)
log_rows.sort(key=lambda r: (r["model"], r["country"], r["case_id"], r["field"]))
CHANGELOG.parent.mkdir(parents=True, exist_ok=True)
with open(CHANGELOG, "w", encoding="utf-8") as f:
for row in log_rows:
f.write(json.dumps(row, ensure_ascii=False) + "\n")
log.info(f"cleaning applied: {n_recovered} recovered, {n_reviewed} reviewed; "
f"{len(log_rows)} changes -> {CHANGELOG}")
def _basic_logging() -> None:
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(message)s",
handlers=[logging.StreamHandler(sys.stderr)],
)
def scan_main() -> None:
_basic_logging()
parser = argparse.ArgumentParser(
prog="legex-refusals-scan",
description="Flag inference cells that need human review (not auto-recoverable).",
)
parser.add_argument("--out", type=Path, default=DEFAULT_XLSX)
parser.add_argument(
"--model", action="append", dest="models", metavar="SLUG",
help="Restrict to this model slug (repeatable). Default: all models.",
)
parser.add_argument(
"--conflicts", type=Path, default=None,
help="Duplicate-run conflict sidecar (JSONL) to fold into the workbook.",
)
args = parser.parse_args()
scan(args.out, set(args.models) if args.models else None, args.conflicts)
def apply_main() -> None:
_basic_logging()
parser = argparse.ArgumentParser(
prog="legex-refusals-apply",
description="Canonicalise recoverable values, apply reviewed decisions, stamp provenance.",
)
parser.add_argument("--xlsx", type=Path, default=DEFAULT_XLSX)
parser.add_argument(
"--model", action="append", dest="models", metavar="SLUG",
help="Restrict to this model slug (repeatable); other models keep their "
"files and changelog history. Default: all models.",
)
args = parser.parse_args()
apply(args.xlsx, set(args.models) if args.models else None)
if __name__ == "__main__":
scan()
|