Spaces:
Sleeping
Sleeping
File size: 15,117 Bytes
5ffc645 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 | #!/usr/bin/env python3
"""
Smoke Signal — Stage 12: Confidence Recalibration
==================================================
Recomputes confidence thresholds from the full gold set using a simple
precision/recall threshold analysis per region class.
Outputs:
- manifest/confidence_calibration.json
- training/runs/<RUN_ID>_recalibration.json
- manifest/run_log.csv entry (governance)
"""
import argparse
import csv
import json
from datetime import datetime, timezone
from pathlib import Path
from typing import Dict, List, Optional, Tuple
ROOT = Path(__file__).resolve().parents[1]
MANIFEST_CSV = ROOT / "manifest" / "source_manifest.csv"
CALIBRATION_JSON = ROOT / "manifest" / "confidence_calibration.json"
RUN_LOG_CSV = ROOT / "manifest" / "run_log.csv"
GOLD_FILE = ROOT / "gold" / "gold_corrections.jsonl"
RUNS_DIR = ROOT / "training" / "runs"
ELIGIBLE_RIGHTS = {"public-domain", "licensed-owned", "controlled-internal"}
BLOCKED_RIGHTS = {"unknown", "excluded"}
RUN_LOG_FIELDS = [
"run_id",
"date",
"operator",
"config_version",
"schema_version",
"source_batch",
"pages_processed",
"errors",
"cost_usd",
"output_path",
"notes",
]
def utc_now() -> datetime:
return datetime.now(timezone.utc)
def utc_iso() -> str:
return utc_now().isoformat().replace("+00:00", "Z")
def ensure_run_dirs() -> None:
RUNS_DIR.mkdir(parents=True, exist_ok=True)
def ensure_run_log() -> None:
RUN_LOG_CSV.parent.mkdir(parents=True, exist_ok=True)
if RUN_LOG_CSV.exists():
return
with open(RUN_LOG_CSV, "w", newline="", encoding="utf-8") as f:
writer = csv.DictWriter(f, fieldnames=RUN_LOG_FIELDS)
writer.writeheader()
def append_run_log(row: Dict[str, str]) -> None:
ensure_run_log()
with open(RUN_LOG_CSV, "a", newline="", encoding="utf-8") as f:
writer = csv.DictWriter(f, fieldnames=RUN_LOG_FIELDS)
writer.writerow({k: row.get(k, "") for k in RUN_LOG_FIELDS})
def load_manifest() -> Dict[str, Dict[str, str]]:
out: Dict[str, Dict[str, str]] = {}
if not MANIFEST_CSV.exists():
return out
with open(MANIFEST_CSV, newline="", encoding="utf-8") as f:
for row in csv.DictReader(f):
book_id = str(row.get("book_id", "")).strip()
if book_id:
out[book_id] = row
return out
def load_gold_records(path: Path) -> List[Dict]:
rows: List[Dict] = []
if not path.exists():
return rows
with open(path, encoding="utf-8") as f:
for idx, line in enumerate(f, start=1):
line = line.strip()
if not line:
continue
try:
rec = json.loads(line)
rec["_line"] = idx
rows.append(rec)
except json.JSONDecodeError:
continue
return rows
def parse_bool(value) -> Optional[bool]:
if isinstance(value, bool):
return value
if value is None:
return None
text = str(value).strip().lower()
if text in {"true", "1", "yes", "y"}:
return True
if text in {"false", "0", "no", "n"}:
return False
return None
def parse_confidence(value) -> Optional[float]:
try:
conf = float(value)
return max(0.0, min(1.0, conf))
except (TypeError, ValueError):
return None
def infer_was_correct(rec: Dict) -> Optional[bool]:
explicit = parse_bool(rec.get("was_correct"))
if explicit is not None:
return explicit
final_text = str(rec.get("final_text", "")).strip()
raw_text = str(rec.get("raw_text", rec.get("raw_ocr", ""))).strip()
if final_text and raw_text:
return final_text == raw_text
return None
def precision_recall_at_threshold(rows: List[Tuple[float, int]], threshold: float) -> Dict[str, float]:
tp = fp = fn = tn = 0
for conf, label in rows:
pred = 1 if conf >= threshold else 0
if pred == 1 and label == 1:
tp += 1
elif pred == 1 and label == 0:
fp += 1
elif pred == 0 and label == 1:
fn += 1
else:
tn += 1
precision = tp / (tp + fp) if (tp + fp) else 0.0
recall = tp / (tp + fn) if (tp + fn) else 0.0
return {
"tp": tp,
"fp": fp,
"fn": fn,
"tn": tn,
"precision": precision,
"recall": recall,
}
def f_beta(precision: float, recall: float, beta: float) -> float:
if precision <= 0 and recall <= 0:
return 0.0
beta2 = beta * beta
denom = (beta2 * precision) + recall
if denom <= 0:
return 0.0
return (1 + beta2) * (precision * recall) / denom
def select_thresholds(
rows: List[Tuple[float, int]],
auto_precision_target: float,
auto_recall_floor: float,
review_recall_target: float,
review_precision_floor: float,
quarantine_gap: float,
) -> Dict:
unique_thresholds = sorted({round(conf, 4) for conf, _ in rows})
if not unique_thresholds:
return {
"auto_accept": 0.85,
"review": 0.60,
"quarantine": 0.35,
"metrics": {},
}
# Add boundary values so we can always compute a fallback.
thresholds = sorted(set([0.0, 1.0] + unique_thresholds))
# Auto-accept: highest threshold meeting strict precision target.
auto_t = None
for t in thresholds:
m = precision_recall_at_threshold(rows, t)
if m["precision"] >= auto_precision_target and m["recall"] >= auto_recall_floor:
auto_t = t
if auto_t is None:
# Fallback: maximize F0.5 to prioritize precision.
auto_t = max(thresholds, key=lambda t: f_beta(
precision_recall_at_threshold(rows, t)["precision"],
precision_recall_at_threshold(rows, t)["recall"],
beta=0.5,
))
# Review threshold: below/at auto threshold, try to capture most true positives.
review_candidates = [t for t in thresholds if t <= auto_t]
review_t = None
for t in review_candidates:
m = precision_recall_at_threshold(rows, t)
if m["recall"] >= review_recall_target and m["precision"] >= review_precision_floor:
review_t = t
break
if review_t is None:
# Fallback: maximize F1 while respecting t <= auto_t.
review_t = max(review_candidates, key=lambda t: f_beta(
precision_recall_at_threshold(rows, t)["precision"],
precision_recall_at_threshold(rows, t)["recall"],
beta=1.0,
))
review_t = min(review_t, auto_t)
quarantine_t = max(0.0, review_t - quarantine_gap)
quarantine_t = min(quarantine_t, review_t)
# Round for readability and stable diffs.
auto_t = round(float(auto_t), 3)
review_t = round(float(review_t), 3)
quarantine_t = round(float(quarantine_t), 3)
# Guarantee monotonic order.
if review_t > auto_t:
review_t = auto_t
if quarantine_t > review_t:
quarantine_t = review_t
return {
"auto_accept": auto_t,
"review": review_t,
"quarantine": quarantine_t,
"metrics": {
"auto": precision_recall_at_threshold(rows, auto_t),
"review": precision_recall_at_threshold(rows, review_t),
"quarantine": precision_recall_at_threshold(rows, quarantine_t),
},
}
def build_region_rows(
gold_records: List[Dict],
manifest: Dict[str, Dict[str, str]],
rights_class_filter: Optional[str],
) -> Tuple[Dict[str, List[Tuple[float, int]]], Dict[str, int]]:
by_region: Dict[str, List[Tuple[float, int]]] = {}
counters = {
"input": len(gold_records),
"used": 0,
"skipped_missing_book": 0,
"skipped_missing_manifest": 0,
"skipped_blocked_rights": 0,
"skipped_rights_filter": 0,
"skipped_missing_confidence": 0,
"skipped_missing_label": 0,
}
for rec in gold_records:
book_id = str(rec.get("book_id", "")).strip()
if not book_id:
counters["skipped_missing_book"] += 1
continue
manifest_row = manifest.get(book_id)
if not manifest_row:
counters["skipped_missing_manifest"] += 1
continue
rights = str(manifest_row.get("rights_class", "unknown")).strip().lower()
if rights in BLOCKED_RIGHTS or rights not in ELIGIBLE_RIGHTS:
counters["skipped_blocked_rights"] += 1
continue
if rights_class_filter and rights != rights_class_filter:
counters["skipped_rights_filter"] += 1
continue
conf = parse_confidence(rec.get("confidence"))
if conf is None:
counters["skipped_missing_confidence"] += 1
continue
was_correct = infer_was_correct(rec)
if was_correct is None:
counters["skipped_missing_label"] += 1
continue
# Positive class = OCR output was correct.
label = 1 if was_correct else 0
region_class = str(rec.get("region_class", "narration")).strip() or "narration"
by_region.setdefault(region_class, []).append((conf, label))
by_region.setdefault("_default", []).append((conf, label))
counters["used"] += 1
return by_region, counters
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Smoke Signal — Stage 12: Recalibration")
parser.add_argument("--gold-file", default=str(GOLD_FILE), help="Path to gold corrections JSONL")
parser.add_argument(
"--rights-class",
default=None,
choices=sorted(ELIGIBLE_RIGHTS),
help="Optional rights class filter",
)
parser.add_argument("--operator", default="codex", help="Operator for governance log")
parser.add_argument("--auto-precision-target", type=float, default=0.98)
parser.add_argument("--auto-recall-floor", type=float, default=0.20)
parser.add_argument("--review-recall-target", type=float, default=0.90)
parser.add_argument("--review-precision-floor", type=float, default=0.60)
parser.add_argument("--quarantine-gap", type=float, default=0.20)
parser.add_argument("--min-samples-per-class", type=int, default=10)
parser.add_argument("--run-id", default=None, help="Optional explicit run id")
return parser.parse_args()
def main() -> None:
args = parse_args()
ensure_run_dirs()
run_id = args.run_id or f"SS-CAL-{utc_now().strftime('%Y%m%d-%H%M%S')}"
gold_path = Path(args.gold_file).expanduser().resolve()
if not gold_path.exists():
raise FileNotFoundError(f"Gold file not found: {gold_path}")
manifest = load_manifest()
if not manifest:
raise RuntimeError("Manifest is empty. Cannot enforce rights controls.")
gold_records = load_gold_records(gold_path)
if not gold_records:
raise RuntimeError("No valid gold records found.")
by_region, counters = build_region_rows(gold_records, manifest, args.rights_class)
if counters["used"] == 0:
raise RuntimeError(f"No usable records for recalibration after filtering. Counters: {counters}")
calibration: Dict[str, Dict] = {}
report_regions: Dict[str, Dict] = {}
for region_class, rows in by_region.items():
if len(rows) < args.min_samples_per_class and region_class != "_default":
# Too little data for a reliable per-class threshold; defer to default.
continue
selected = select_thresholds(
rows=rows,
auto_precision_target=args.auto_precision_target,
auto_recall_floor=args.auto_recall_floor,
review_recall_target=args.review_recall_target,
review_precision_floor=args.review_precision_floor,
quarantine_gap=args.quarantine_gap,
)
corrections = sum(1 for _, label in rows if label == 0)
calibration[region_class] = {
"auto_accept": selected["auto_accept"],
"review": selected["review"],
"quarantine": selected["quarantine"],
"corrections": corrections,
}
report_regions[region_class] = {
"samples": len(rows),
"correct": sum(1 for _, label in rows if label == 1),
"incorrect": sum(1 for _, label in rows if label == 0),
"thresholds": calibration[region_class],
"metrics": selected["metrics"],
}
# Ensure required defaults exist for runtime readers.
if "_default" not in calibration:
calibration["_default"] = {
"auto_accept": 0.85,
"review": 0.60,
"quarantine": 0.35,
"corrections": 0,
}
default_entry = calibration["_default"]
for cls in ["narration", "dialogue-speech-bubble", "caption", "title", "sign-label"]:
if cls not in calibration:
calibration[cls] = dict(default_entry)
CALIBRATION_JSON.parent.mkdir(parents=True, exist_ok=True)
with open(CALIBRATION_JSON, "w", encoding="utf-8") as f:
json.dump(calibration, f, indent=2, ensure_ascii=False)
report = {
"run_id": run_id,
"generated_at": utc_iso(),
"config_version": "ss_confidence_calibration_v0.1",
"schema_version": "ss_confidence_calibration_report_v1",
"gold_file": str(gold_path),
"rights_class_filter": args.rights_class,
"counters": counters,
"regions": report_regions,
"output_file": str(CALIBRATION_JSON),
}
report_path = RUNS_DIR / f"{run_id}_recalibration.json"
with open(report_path, "w", encoding="utf-8") as f:
json.dump(report, f, indent=2, ensure_ascii=False)
append_run_log(
{
"run_id": run_id,
"date": utc_now().date().isoformat(),
"operator": args.operator,
"config_version": "ss_confidence_calibration_v0.1",
"schema_version": "ss_confidence_calibration_report_v1",
"source_batch": args.rights_class or "auto",
"pages_processed": str(counters["used"]),
"errors": str(
counters["skipped_missing_book"]
+ counters["skipped_missing_manifest"]
+ counters["skipped_blocked_rights"]
+ counters["skipped_rights_filter"]
+ counters["skipped_missing_confidence"]
+ counters["skipped_missing_label"]
),
"cost_usd": "",
"output_path": str(CALIBRATION_JSON.relative_to(ROOT)),
"notes": json.dumps(
{
"auto_precision_target": args.auto_precision_target,
"review_recall_target": args.review_recall_target,
"regions_calibrated": sorted(report_regions.keys()),
},
ensure_ascii=False,
),
}
)
print(f"Recalibration complete: {CALIBRATION_JSON}")
print(f"Report: {report_path}")
print(f"Governance log updated: {RUN_LOG_CSV}")
if __name__ == "__main__":
main()
|