File size: 16,148 Bytes
87cc7f2 | 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 | """Conservative final-answer extraction and scoring for the MATH-500 pilot.
Scoring is deliberately three-way at its core: a confidently parsed answer can
be marked correct or incorrect; answers that cannot be interpreted reliably
remain unparsed. Completion extraction can additionally abstain when the output
contains a malformed or explicitly ambiguous final answer.
``math_verify`` is used when installed, but runs in a child process so a hard
wall-clock limit applies even on platforms where its own signal-based timeout
is unavailable. Without it, the fallback only compares identical normalized
answers or exact numeric scalars/fractions.
"""
from __future__ import annotations
import importlib.util
import json
import re
import subprocess
import sys
from dataclasses import dataclass
from decimal import Decimal, InvalidOperation
from fractions import Fraction
from typing import Literal
ExtractionStatus = Literal["parsed", "abstained", "unparsed"]
ScoreStatus = Literal["correct", "incorrect", "abstained", "unparsed"]
@dataclass(frozen=True)
class ExtractionResult:
status: ExtractionStatus
answer: str | None = None
reason: str = ""
@dataclass(frozen=True)
class ScoreResult:
status: ScoreStatus
predicted_answer: str | None
expected_answer: str
method: str
reason: str = ""
_BOX_OPEN = re.compile(r"\\(?:boxed|fbox)\s*\{")
_FINAL_MARKER = re.compile(
r"(?:\bfinal\s+answer\b|\banswer\b)\s*(?:is\s*)?(?:[:=]\s*)?",
re.IGNORECASE,
)
_REASONING_AFTER_ANSWER = re.compile(
r"\b(?:because|since|therefore|thus|so|which\s+means|this\s+gives)\b",
re.IGNORECASE,
)
_MATH_VERIFY_WORKER = r'''
import json, sys
predicted, expected = json.load(sys.stdin)
try:
from math_verify import ExprExtractionConfig, LatexExtractionConfig, parse, verify
except ImportError:
print(json.dumps({"kind": "unavailable"}))
raise SystemExit(0)
try:
extraction = [LatexExtractionConfig(), ExprExtractionConfig()]
parse_options = {
"extraction_config": extraction,
"fallback_mode": "no_fallback",
"extraction_mode": "first_match",
"parsing_timeout": 0,
"raise_on_error": True,
}
parsed_expected = parse(expected, **parse_options)
parsed_predicted = parse(predicted, **parse_options)
if not parsed_expected or not parsed_predicted:
print(json.dumps({"kind": "unparsed"}))
else:
matches = verify(
parsed_expected,
parsed_predicted,
timeout_seconds=0,
raise_on_error=True,
)
print(json.dumps({"kind": "verified", "matches": bool(matches)}))
except BaseException as exc:
print(json.dumps({"kind": "error", "error": type(exc).__name__}))
'''
def _read_braced(text: str, open_brace: int) -> tuple[str | None, int]:
"""Read a balanced brace expression; return its contents and end offset."""
depth = 0
escaped = False
for position in range(open_brace, len(text)):
char = text[position]
if escaped:
escaped = False
continue
if char == "\\":
escaped = True
continue
if char == "{":
depth += 1
elif char == "}":
depth -= 1
if depth == 0:
return text[open_brace + 1 : position], position + 1
return None, len(text)
def _boxed_answers(text: str) -> tuple[list[tuple[str, int, int]], bool]:
found: list[tuple[str, int, int]] = []
malformed = False
for match in _BOX_OPEN.finditer(text):
open_brace = match.end() - 1
answer, end = _read_braced(text, open_brace)
if answer is None:
malformed = True
continue
found.append((answer.strip(), match.start(), end))
return found, malformed
def _clean_final_candidate(candidate: str) -> str:
candidate = candidate.strip()
candidate = re.sub(r"^\*{1,2}|\*{1,2}$", "", candidate).strip()
# A single terminal sentence mark is prose punctuation, not part of a value.
candidate = re.sub(r"[.!?]+\s*$", "", candidate).strip()
return candidate
def _answer_from_marker(text: str) -> tuple[str | None, int | None, bool]:
"""Get the tail after the last explicit answer label, if it is self-contained."""
matches = list(_FINAL_MARKER.finditer(text))
if not matches:
return None, None, False
marker = matches[-1]
candidate = _clean_final_candidate(text[marker.end() :])
if not candidate:
return None, marker.start(), True
if "\n" in candidate or _REASONING_AFTER_ANSWER.search(candidate):
return None, marker.start(), True
# Avoid treating a sentence that continues after the answer as a clean value.
if len(candidate.split()) > 12:
return None, marker.start(), True
return candidate, marker.start(), False
def extract_final_answer(completion: str) -> ExtractionResult:
"""Extract an explicitly marked final answer without guessing from reasoning.
Accepted forms are a final ``\\boxed{...}``/``\\fbox{...}``, a terminal
``answer: ...`` or ``final answer is ...`` label, or a terminal ``#### ...``
line. A bare number or expression without one of these markers is unparsed.
"""
if not isinstance(completion, str) or not completion.strip():
return ExtractionResult("unparsed", reason="empty or non-text completion")
boxes, malformed_box = _boxed_answers(completion)
boxed: tuple[str, int, int] | None = boxes[-1] if boxes else None
marker_answer, marker_position, malformed_marker = _answer_from_marker(completion)
if malformed_box and (boxed is None or completion.rfind(r"\boxed") > boxed[1]):
return ExtractionResult("abstained", reason="final boxed answer is incomplete")
if malformed_marker and (marker_position is not None) and (
boxed is None or marker_position >= boxed[1]
):
return ExtractionResult("abstained", reason="answer label is ambiguous or followed by explanation")
# A later explicit answer label takes precedence over an earlier boxed value.
if marker_answer is not None and (boxed is None or marker_position >= boxed[1]):
if re.search(r"\b(?:or|either)\b", marker_answer, re.IGNORECASE):
return ExtractionResult("abstained", reason="final answer presents alternatives")
return ExtractionResult("parsed", marker_answer, "terminal answer label")
if boxed is not None:
answer, _, end = boxed
if not answer:
return ExtractionResult("abstained", reason="final boxed answer is empty")
# In a construction such as ``\\boxed{1} or \\boxed{2}``, selecting one
# answer would be arbitrary. Earlier boxed work by itself is not ambiguous.
if len(boxes) > 1 and re.search(
r"\b(?:or|either)\b", completion[boxes[-2][2] : boxes[-1][1]], re.IGNORECASE
):
return ExtractionResult("abstained", reason="final answer presents alternatives")
if re.search(r"\b(?:or|either)\b", completion[end : end + 80], re.IGNORECASE):
return ExtractionResult("abstained", reason="boxed answer is followed by an alternative")
return ExtractionResult("parsed", answer, "last complete boxed answer")
# Support the common GSM-style final answer marker for generated pilot text.
final_lines = re.findall(r"(?m)^\s*####\s*(.*?)\s*$", completion)
if final_lines:
answer = _clean_final_candidate(final_lines[-1])
if answer and not re.search(r"\b(?:or|either)\b", answer, re.IGNORECASE):
return ExtractionResult("parsed", answer, "terminal #### answer line")
return ExtractionResult("abstained", reason="final answer line is empty or ambiguous")
return ExtractionResult("unparsed", reason="no explicit final-answer marker found")
def _unwrap_math_delimiters(value: str) -> str:
value = value.strip().replace("−", "-").replace("+", "+").replace("×", "*").replace("÷", "/")
value = value.replace(r"\left", "").replace(r"\right", "")
pairs = (("$", "$"), (r"\(", r"\)"), (r"\[", r"\]"))
changed = True
while changed:
changed = False
for left, right in pairs:
if value.startswith(left) and value.endswith(right) and len(value) >= len(left) + len(right):
value = value[len(left) : len(value) - len(right)].strip()
changed = True
return value
_NUMBER = r"[+-]?(?:\d+(?:\.\d*)?|\.\d+)"
_PLAIN_FRACTION = re.compile(rf"^\s*({_NUMBER})\s*/\s*({_NUMBER})\s*$")
_LATEX_FRACTION = re.compile(rf"^\\(?:dfrac|tfrac|frac)\s*\{{\s*({_NUMBER})\s*\}}\s*\{{\s*({_NUMBER})\s*\}}$")
_SIMPLE_NUMBER = re.compile(rf"^\s*({_NUMBER})\s*$")
def _fraction_from_decimal(text: str) -> Fraction:
if len(text) > 256:
raise ValueError("numeric answer is too long")
return Fraction(Decimal(text))
def _numeric_value(value: str) -> Fraction | None:
value = _unwrap_math_delimiters(value)
if len(value) > 512:
return None
match = _LATEX_FRACTION.fullmatch(value)
if match:
try:
return _fraction_from_decimal(match.group(1)) / _fraction_from_decimal(match.group(2))
except (InvalidOperation, ValueError, ZeroDivisionError):
return None
match = _PLAIN_FRACTION.fullmatch(value)
if match:
try:
return _fraction_from_decimal(match.group(1)) / _fraction_from_decimal(match.group(2))
except (InvalidOperation, ValueError, ZeroDivisionError):
return None
match = _SIMPLE_NUMBER.fullmatch(value)
if match:
try:
return _fraction_from_decimal(match.group(1))
except (InvalidOperation, ValueError, ZeroDivisionError):
return None
return None
def _canonical_literal(value: str) -> str:
value = _unwrap_math_delimiters(value)
value = re.sub(r"\s+", "", value)
return value
def _fallback_score(predicted: str, expected: str) -> tuple[ScoreStatus, str, str]:
predicted_number = _numeric_value(predicted)
expected_number = _numeric_value(expected)
if predicted_number is not None and expected_number is not None:
if predicted_number == expected_number:
return "correct", "exact_numeric", "exact numeric values match"
return "incorrect", "exact_numeric", "exact numeric values differ"
if _canonical_literal(predicted) and _canonical_literal(predicted) == _canonical_literal(expected):
return "correct", "exact_literal", "normalized answer text matches exactly"
return "unparsed", "none", "unsupported answer form without math_verify"
def _math_verify_available() -> bool:
try:
return importlib.util.find_spec("math_verify") is not None
except (ImportError, ValueError):
return False
def _run_math_verify(predicted: str, expected: str, timeout_seconds: float) -> tuple[str, str] | None:
if not _math_verify_available():
return None
try:
process = subprocess.Popen(
[sys.executable, "-c", _MATH_VERIFY_WORKER],
stdin=subprocess.PIPE,
stdout=subprocess.PIPE,
stderr=subprocess.DEVNULL,
text=True,
)
try:
stdout, _ = process.communicate(
input=json.dumps([predicted, expected]), timeout=timeout_seconds
)
except subprocess.TimeoutExpired:
process.kill()
process.communicate()
return "timeout", "math_verify exceeded the scoring timeout"
if process.returncode != 0 or not stdout.strip():
return "error", "math_verify subprocess failed"
payload = json.loads(stdout.strip().splitlines()[-1])
except (OSError, ValueError, json.JSONDecodeError):
return "error", "math_verify subprocess could not be read"
kind = payload.get("kind")
if kind == "unavailable":
return None
if kind == "verified":
return ("correct" if payload.get("matches") else "incorrect"), "math_verify equivalence check"
if kind == "unparsed":
return "unparsed", "math_verify could not parse one or both answers"
if kind == "error":
return "error", f"math_verify raised {payload.get('error', 'an error')}"
return "error", "math_verify returned an unexpected result"
def score_extracted_answer(
predicted_answer: str,
expected_answer: str,
*,
timeout_seconds: float = 3.0,
) -> ScoreResult:
"""Score two answer strings, abstaining whenever the comparison is unclear."""
if not isinstance(predicted_answer, str) or not predicted_answer.strip():
return ScoreResult("unparsed", None, str(expected_answer), "none", "prediction is empty or non-text")
if not isinstance(expected_answer, str) or not expected_answer.strip():
return ScoreResult("unparsed", predicted_answer, str(expected_answer), "none", "reference answer is empty or non-text")
if timeout_seconds <= 0:
raise ValueError("timeout_seconds must be positive")
verified = _run_math_verify(predicted_answer, expected_answer, timeout_seconds)
if verified is not None:
status, reason = verified
if status == "correct":
return ScoreResult("correct", predicted_answer, expected_answer, "math_verify", reason)
if status == "incorrect":
return ScoreResult("incorrect", predicted_answer, expected_answer, "math_verify", reason)
if status == "timeout":
return ScoreResult("unparsed", predicted_answer, expected_answer, "math_verify", reason)
status, method, reason = _fallback_score(predicted_answer, expected_answer)
if verified is not None and reason == "unsupported answer form without math_verify":
reason = f"{verified[1]}; {reason}"
return ScoreResult(status, predicted_answer, expected_answer, method, reason)
def score_completion(
completion: str,
expected_answer: str,
*,
timeout_seconds: float = 3.0,
) -> ScoreResult:
"""Extract and score a generated completion against its reference answer."""
extracted = extract_final_answer(completion)
if extracted.status == "unparsed":
return ScoreResult("unparsed", None, str(expected_answer), "none", extracted.reason)
if extracted.status == "abstained":
return ScoreResult("abstained", None, str(expected_answer), "none", extracted.reason)
result = score_extracted_answer(
extracted.answer or "", expected_answer, timeout_seconds=timeout_seconds
)
return ScoreResult(
result.status,
result.predicted_answer,
result.expected_answer,
result.method,
result.reason,
)
def score_answer(
text: str,
reference: str,
*,
timeout_seconds: float = 3.0,
) -> dict[str, object]:
"""Return a JSON-serializable scoring record for a generated completion.
``correct`` is ``True`` for a verified match, ``False`` for a confidently
verified mismatch, and ``None`` when extraction or comparison is uncertain.
"""
result = score_completion(text, reference, timeout_seconds=timeout_seconds)
correct: bool | None
if result.status == "correct":
correct = True
elif result.status == "incorrect":
correct = False
else:
correct = None
return {
"correct": correct,
"status": result.status,
"answer": result.predicted_answer,
"reference": result.expected_answer,
"method": result.method,
"reason": result.reason,
}
__all__ = [
"ExtractionResult",
"ScoreResult",
"extract_final_answer",
"score_answer",
"score_completion",
"score_extracted_answer",
]
|