File size: 18,000 Bytes
fbc14e7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 | """
Token threshold calibration analysis.
Collects evidence quality metrics paired with HHEM scores to find
the optimal MIN_EVIDENCE_TOKENS threshold.
Usage:
python scripts/calibrate_token_threshold.py --samples 200
python scripts/calibrate_token_threshold.py --analyze-only # Re-analyze existing data
"""
from __future__ import annotations
import argparse
import json
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import TYPE_CHECKING
import numpy as np
from sage.config import (
CHARS_PER_TOKEN,
EVALUATION_QUERIES,
MIN_EVIDENCE_CHUNKS,
MIN_EVIDENCE_TOKENS,
get_logger,
log_banner,
log_section,
)
from sage.core import AggregationMethod, ProductScore
from sage.core.evidence import check_evidence_quality
from sage.services.retrieval import get_candidates
if TYPE_CHECKING:
from sage.adapters.hhem import HallucinationDetector
from sage.services.explanation import Explainer
logger = get_logger(__name__)
DEFAULT_OUTPUT = Path("data/calibration/token_threshold.json")
@dataclass
class CalibrationRecord:
"""Single data point for threshold calibration."""
query: str
product_id: str
# Evidence metrics
chunk_count: int
total_tokens: int
min_chunk_tokens: int
max_chunk_tokens: int
top_score: float
avg_rating: float
# Outcome metrics
gate_passed: bool
hhem_score: float | None
is_hallucinated: bool | None
def estimate_tokens(text: str) -> int:
"""Estimate token count from character count."""
return len(text) // CHARS_PER_TOKEN
def collect_evidence_metrics(product: ProductScore) -> dict:
"""Extract detailed evidence metrics for a product."""
chunks = product.evidence
if not chunks:
return {
"chunk_count": 0,
"total_tokens": 0,
"min_chunk_tokens": 0,
"max_chunk_tokens": 0,
}
chunk_tokens = [estimate_tokens(c.text) for c in chunks]
return {
"chunk_count": len(chunks),
"total_tokens": sum(chunk_tokens),
"min_chunk_tokens": min(chunk_tokens),
"max_chunk_tokens": max(chunk_tokens),
}
def collect_calibration_data(
explainer: Explainer,
detector: HallucinationDetector,
max_samples: int = 200,
output_path: Path | None = None,
) -> list[CalibrationRecord]:
"""
Collect calibration data by running pipeline with quality gate disabled.
Key insight: We disable the gate to collect data on explanations that
WOULD have been refused, so we can see their actual HHEM scores.
"""
records: list[CalibrationRecord] = []
log_banner(logger, "COLLECTING CALIBRATION DATA", width=60)
logger.info("Target samples: %d", max_samples)
for query in EVALUATION_QUERIES:
# Get more candidates than usual to capture edge cases
products = get_candidates(
query=query,
k=10,
min_rating=2.0, # Lower threshold to get thin-evidence cases
aggregation=AggregationMethod.MAX,
)
for product in products:
if len(records) >= max_samples:
break
# Check what the gate WOULD have decided
quality = check_evidence_quality(product)
evidence_metrics = collect_evidence_metrics(product)
# Generate explanation WITH GATE DISABLED to see actual outcome
result = explainer.generate_explanation(
query,
product,
max_evidence=5,
enforce_quality_gate=False,
)
# Measure faithfulness
hhem = detector.check_explanation(
result.evidence_texts,
result.explanation,
)
record = CalibrationRecord(
query=query,
product_id=product.product_id,
chunk_count=evidence_metrics["chunk_count"],
total_tokens=evidence_metrics["total_tokens"],
min_chunk_tokens=evidence_metrics["min_chunk_tokens"],
max_chunk_tokens=evidence_metrics["max_chunk_tokens"],
top_score=quality.top_score,
avg_rating=product.avg_rating,
gate_passed=quality.is_sufficient,
hhem_score=hhem.score,
is_hallucinated=hhem.is_hallucinated,
)
records.append(record)
logger.info(
"Sample %d: tokens=%d, gate=%s, hhem=%.3f",
len(records),
record.total_tokens,
"PASS" if record.gate_passed else "FAIL",
record.hhem_score,
)
if len(records) >= max_samples:
break
# Save raw data
if output_path:
output_path.parent.mkdir(parents=True, exist_ok=True)
with open(output_path, "w") as f:
json.dump([asdict(r) for r in records], f, indent=2)
logger.info("Saved %d records to %s", len(records), output_path)
return records
def load_calibration_data(path: Path) -> list[CalibrationRecord]:
"""Load previously collected calibration data."""
with open(path) as f:
data = json.load(f)
return [CalibrationRecord(**r) for r in data]
def _compute_classification_metrics(
would_pass: np.ndarray, hallucinated: np.ndarray
) -> dict:
"""Compute precision, recall, F1 for a gate decision."""
tp = int(np.sum(would_pass & ~hallucinated))
fp = int(np.sum(would_pass & hallucinated))
tn = int(np.sum(~would_pass & hallucinated))
fn = int(np.sum(~would_pass & ~hallucinated))
precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
f1 = (
2 * precision * recall / (precision + recall)
if (precision + recall) > 0
else 0.0
)
return {
"precision": round(precision, 3),
"recall": round(recall, 3),
"f1": round(f1, 3),
"tp": tp,
"fp": fp,
"tn": tn,
"fn": fn,
"n_passed": int(would_pass.sum()),
"n_refused": int((~would_pass).sum()),
}
@dataclass
class _ValidatedRecords:
"""Pre-validated records with extracted numpy arrays."""
tokens: np.ndarray
chunks: np.ndarray
hhem: np.ndarray
hallucinated: np.ndarray
count: int
def _validate_records(
records: list[CalibrationRecord], min_samples: int = 10
) -> _ValidatedRecords | None:
"""Filter and validate records, returning None if insufficient."""
valid = [r for r in records if r.hhem_score is not None]
if len(valid) < min_samples:
logger.warning("Not enough valid samples for analysis: %d", len(valid))
return None
return _ValidatedRecords(
tokens=np.array([r.total_tokens for r in valid]),
chunks=np.array([r.chunk_count for r in valid]),
hhem=np.array([r.hhem_score for r in valid]),
hallucinated=np.array([r.is_hallucinated for r in valid]),
count=len(valid),
)
def _sweep_threshold(
values: np.ndarray,
hallucinated: np.ndarray,
hhem: np.ndarray,
thresholds: list[int],
) -> list[dict]:
"""Sweep a single threshold dimension and compute metrics."""
results = []
for thresh in thresholds:
would_pass = values >= thresh
metrics = _compute_classification_metrics(would_pass, hallucinated)
metrics["threshold"] = thresh
metrics["mean_hhem_above"] = (
round(float(hhem[would_pass].mean()), 3) if would_pass.any() else 0.0
)
results.append(metrics)
return results
def analyze_threshold(records: list[CalibrationRecord]) -> dict:
"""Analyze data to find optimal token threshold."""
validated = _validate_records(records)
if validated is None:
return {"error": "insufficient_samples", "count": len(records)}
thresholds = list(range(20, 200, 10))
results = _sweep_threshold(
validated.tokens, validated.hallucinated, validated.hhem, thresholds
)
best = max(results, key=lambda x: x["f1"])
correlation = float(np.corrcoef(validated.tokens, validated.hhem)[0, 1])
return {
"recommended_threshold": best["threshold"],
"best_f1": best["f1"],
"best_precision": best["precision"],
"best_recall": best["recall"],
"correlation_tokens_hhem": round(correlation, 3),
"current_threshold": MIN_EVIDENCE_TOKENS,
"total_samples": validated.count,
"hallucination_rate": round(float(validated.hallucinated.mean()), 3),
"mean_tokens": round(float(validated.tokens.mean()), 1),
"median_tokens": round(float(np.median(validated.tokens)), 1),
"all_thresholds": results,
}
def analyze_chunk_threshold(records: list[CalibrationRecord]) -> dict:
"""Analyze optimal MIN_EVIDENCE_CHUNKS threshold."""
validated = _validate_records(records)
if validated is None:
return {"error": "insufficient_samples", "count": len(records)}
thresholds = [1, 2, 3, 4, 5]
results = _sweep_threshold(
validated.chunks, validated.hallucinated, validated.hhem, thresholds
)
best = max(results, key=lambda x: x["f1"])
correlation = float(np.corrcoef(validated.chunks, validated.hhem)[0, 1])
return {
"recommended_threshold": best["threshold"],
"best_f1": best["f1"],
"best_precision": best["precision"],
"best_recall": best["recall"],
"correlation_chunks_hhem": round(correlation, 3),
"current_threshold": MIN_EVIDENCE_CHUNKS,
"mean_chunks": round(float(validated.chunks.mean()), 2),
"all_thresholds": results,
}
def analyze_combined_thresholds(records: list[CalibrationRecord]) -> dict:
"""2D sweep of token and chunk thresholds to find optimal combination."""
validated = _validate_records(records)
if validated is None:
return {"error": "insufficient_samples", "count": len(records)}
token_thresholds = [20, 30, 40, 50, 75, 100]
chunk_thresholds = [1, 2, 3]
results = []
for tok_thresh in token_thresholds:
for chunk_thresh in chunk_thresholds:
would_pass = (validated.tokens >= tok_thresh) & (
validated.chunks >= chunk_thresh
)
metrics = _compute_classification_metrics(
would_pass, validated.hallucinated
)
metrics["token_threshold"] = tok_thresh
metrics["chunk_threshold"] = chunk_thresh
results.append(metrics)
best = max(results, key=lambda x: x["f1"])
return {
"recommended_token_threshold": best["token_threshold"],
"recommended_chunk_threshold": best["chunk_threshold"],
"best_f1": best["f1"],
"best_precision": best["precision"],
"best_recall": best["recall"],
"current_token_threshold": MIN_EVIDENCE_TOKENS,
"current_chunk_threshold": MIN_EVIDENCE_CHUNKS,
"all_combinations": results,
}
def print_token_analysis(analysis: dict) -> None:
"""Print token threshold analysis."""
if "error" in analysis:
logger.error("Analysis failed: %s", analysis["error"])
return
log_banner(logger, "TOKEN THRESHOLD ANALYSIS", width=60)
log_section(logger, "Summary")
logger.info("Total samples: %d", analysis["total_samples"])
logger.info("Mean tokens: %.1f", analysis["mean_tokens"])
logger.info("Median tokens: %.1f", analysis["median_tokens"])
logger.info("Hallucination rate: %.1f%%", analysis["hallucination_rate"] * 100)
logger.info("Token-HHEM corr: %+.3f", analysis["correlation_tokens_hhem"])
log_section(logger, "Recommendation")
logger.info("Current threshold: %d tokens", analysis["current_threshold"])
logger.info("Recommended: %d tokens", analysis["recommended_threshold"])
logger.info("Best F1: %.3f", analysis["best_f1"])
logger.info("Best precision: %.3f", analysis["best_precision"])
logger.info("Best recall: %.3f", analysis["best_recall"])
log_section(logger, "Token Threshold Sweep")
logger.info("thresh | F1 | prec | recall | passed | refused")
logger.info("-------|-------|-------|--------|--------|--------")
for r in analysis["all_thresholds"]:
marker = " *" if r["threshold"] == analysis["recommended_threshold"] else ""
logger.info(
" %3d | %.3f | %.3f | %.3f | %3d | %3d%s",
r["threshold"],
r["f1"],
r["precision"],
r["recall"],
r["n_passed"],
r["n_refused"],
marker,
)
def print_chunk_analysis(analysis: dict) -> None:
"""Print chunk threshold analysis."""
if "error" in analysis:
logger.error("Chunk analysis failed: %s", analysis["error"])
return
log_banner(logger, "CHUNK THRESHOLD ANALYSIS", width=60)
log_section(logger, "Summary")
logger.info("Mean chunks: %.2f", analysis["mean_chunks"])
logger.info("Chunk-HHEM corr: %+.3f", analysis["correlation_chunks_hhem"])
log_section(logger, "Recommendation")
logger.info("Current threshold: %d chunks", analysis["current_threshold"])
logger.info("Recommended: %d chunk(s)", analysis["recommended_threshold"])
logger.info("Best F1: %.3f", analysis["best_f1"])
log_section(logger, "Chunk Threshold Sweep")
logger.info("chunks | F1 | prec | recall | passed | refused")
logger.info("-------|-------|-------|--------|--------|--------")
for r in analysis["all_thresholds"]:
marker = " *" if r["threshold"] == analysis["recommended_threshold"] else ""
logger.info(
" %d | %.3f | %.3f | %.3f | %3d | %3d%s",
r["threshold"],
r["f1"],
r["precision"],
r["recall"],
r["n_passed"],
r["n_refused"],
marker,
)
def print_combined_analysis(analysis: dict) -> None:
"""Print combined threshold analysis."""
if "error" in analysis:
logger.error("Combined analysis failed: %s", analysis["error"])
return
log_banner(logger, "COMBINED THRESHOLD ANALYSIS", width=60)
log_section(logger, "Current vs Recommended")
logger.info(
"Current: %d tokens, %d chunks",
analysis["current_token_threshold"],
analysis["current_chunk_threshold"],
)
logger.info(
"Recommended: %d tokens, %d chunk(s)",
analysis["recommended_token_threshold"],
analysis["recommended_chunk_threshold"],
)
logger.info("Best F1: %.3f", analysis["best_f1"])
logger.info("Precision: %.3f", analysis["best_precision"])
logger.info("Recall: %.3f", analysis["best_recall"])
log_section(logger, "2D Threshold Grid")
logger.info("tokens | chunks | F1 | prec | recall | passed | refused")
logger.info("-------|--------|-------|-------|--------|--------|--------")
for r in sorted(
analysis["all_combinations"],
key=lambda x: (x["token_threshold"], x["chunk_threshold"]),
):
is_best = (
r["token_threshold"] == analysis["recommended_token_threshold"]
and r["chunk_threshold"] == analysis["recommended_chunk_threshold"]
)
is_current = (
r["token_threshold"] == analysis["current_token_threshold"]
and r["chunk_threshold"] == analysis["current_chunk_threshold"]
)
marker = " *" if is_best else (" (current)" if is_current else "")
logger.info(
" %3d | %d | %.3f | %.3f | %.3f | %3d | %3d%s",
r["token_threshold"],
r["chunk_threshold"],
r["f1"],
r["precision"],
r["recall"],
r["n_passed"],
r["n_refused"],
marker,
)
def main() -> None:
parser = argparse.ArgumentParser(
description="Calibrate MIN_EVIDENCE_TOKENS and MIN_EVIDENCE_CHUNKS thresholds"
)
parser.add_argument(
"--samples",
type=int,
default=200,
help="Number of samples to collect",
)
parser.add_argument(
"--output",
type=Path,
default=DEFAULT_OUTPUT,
help="Output path for calibration data",
)
parser.add_argument(
"--analyze-only",
action="store_true",
help="Skip collection, analyze existing data",
)
args = parser.parse_args()
if args.analyze_only:
if not args.output.exists():
logger.error("No data file found at %s", args.output)
return
records = load_calibration_data(args.output)
logger.info("Loaded %d records from %s", len(records), args.output)
else:
from sage.adapters.hhem import HallucinationDetector
from sage.services.explanation import Explainer
explainer = Explainer()
detector = HallucinationDetector()
records = collect_calibration_data(
explainer, detector, args.samples, args.output
)
# Run all three analyses
token_analysis = analyze_threshold(records)
chunk_analysis = analyze_chunk_threshold(records)
combined_analysis = analyze_combined_thresholds(records)
# Print results
print_token_analysis(token_analysis)
print_chunk_analysis(chunk_analysis)
print_combined_analysis(combined_analysis)
# Save all analyses
all_analysis = {
"token_analysis": token_analysis,
"chunk_analysis": chunk_analysis,
"combined_analysis": combined_analysis,
}
analysis_path = args.output.with_suffix(".analysis.json")
with open(analysis_path, "w") as f:
json.dump(all_analysis, f, indent=2)
logger.info("Saved analysis to %s", analysis_path)
if __name__ == "__main__":
main()
|