File size: 27,739 Bytes
a512fa8 6646976 a512fa8 799fa1f a512fa8 | 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 | from __future__ import annotations
from collections import Counter, defaultdict
import contextlib
import functools
import importlib.util
import io
import json
import math
from multiprocessing.managers import BaseManager, BaseProxy
from pathlib import Path
import re
import shlex
import statistics
import sys
import textwrap
import threading
from typing import Any, Iterable, Mapping
import numpy as np
from scoring.protocols import PYTHON_FENCE_CONTRACT, code_contract_status, mechanical_repetitive_tail_metrics, think_protocol_status
from scoring.verification import CPU_VERIFIER_REVISION, code_blocks, compare_answers, deterministic_verifier_contract, typed_declared_answer
from evaluation.serve_verifier import ExternalVerifier, VerifierEngine, prepare_semantic_record, record_visible_response, missing_answer_outcome
OFFICIAL_RUNNERS = Path(__file__).resolve().parents[1] / 'third_party' / 'official_evaluators'
MACRO_AVERAGE_SUBSET_FIELDS = {'BBH': 'task', 'BBH128': 'task', 'BBEH': 'task', 'CMMLU': 'subject', 'CMMLU128': 'subject'}
ACCEPTED_CODE_LANGUAGES = {'', 'py', 'python', 'python3'}
PYTHON_OPEN = re.compile('```[ \\t]*(?:python|py|python3)[ \\t]*\\r?\\n', re.I)
TRAILING_IM_END = re.compile('(?:<\\|im_end\\|>)?\\s*$')
MATH_SOURCES = {'AIME2024', 'AIME2025', 'AIME2026', 'HMMT2025', 'MATH500', 'HMMT2026'}
CHOICE_SOURCES = {'GPQA68', 'MMLU256', 'HellaSwag128', 'WinoGrande128', 'ARC-C128', 'ARC-E128', 'CEval128', 'CMMLU128', 'MMLUPro', 'MMLURedux', 'SuperGPQA'}
CODE_SOURCES = {'HumanEval', 'MBPP128', 'LiveCodeBenchV6'}
_BBEH_LOAD_LOCK = threading.Lock()
def evaluation_finish_reason(row: Mapping[str, Any]) -> str:
return str(row.get('finish_reason') or ('eos' if row.get('finished_by_eos') else 'max_tokens')).strip().lower()
def references(row: Mapping[str, Any], metadata: Mapping[str, Any]) -> list[Any]:
alternatives = metadata.get('reference_answers')
if isinstance(alternatives, list) and alternatives:
return alternatives
return [row['reference_answer']]
def adapt_score_row(row):
adapted = {**row, 'source': row['score_source']}
if row['score_source'] == 'HellaSwag128':
adapted['reference_answer'] = str(ord(row['reference_answer']) - 65)
elif row['score_source'] == 'WinoGrande128':
adapted['reference_answer'] = str(ord(row['reference_answer']) - 64)
return adapted
def choice_references(source: str, values: Iterable[Any]) -> list[str]:
if source == 'HellaSwag128':
return [chr(ord('A') + int(value)) for value in values]
if source == 'WinoGrande128':
return [chr(ord('A') + int(value) - 1) for value in values]
return [str(value).strip().upper() for value in values]
def math_contract(row: Mapping[str, Any], values: list[Any]) -> dict[str, Any]:
extra = {'verifier_mode': 'math_equivalence', 'problem': str(row['prompt'])}
if len(values) > 1:
extra['reference_answers'] = values
routed = deterministic_verifier_contract({'task_type': 'math', 'ground_truth': values[0], 'prompt': [{'role': 'user', 'content': str(row['prompt'])}], 'extra_info': extra})
if routed['eligible']:
return routed
return {'eligible': True, 'reason': routed['reason'], 'verifier_mode': 'math_equivalence', 'answer_type': 'symbolic', 'resolution': 'math_verify_then_symbolic'}
def standard_verifier_record(row: Mapping[str, Any], response: str, *, accept_unmarked_final: bool=False) -> tuple[dict[str, Any], dict[str, Any]]:
source = str(row['source'])
metadata = json.loads(str(row.get('metadata_json') or '{}'))
values = references(row, metadata)
finish_reason = evaluation_finish_reason(row)
record: dict[str, Any] = {'sample_id': f"{row['checkpoint_id']}:{row['source_sample_id']}:{row['sample_index']}", 'problem': str(row['prompt']), 'response': response, 'finish_reason': finish_reason, 'reference_answer': values[0]}
if len(values) > 1:
record['reference_answers'] = values
routing: dict[str, Any]
if source in MATH_SOURCES:
routing = math_contract(row, values)
record.update(task_type='math', verifier_mode=routing['verifier_mode'], answer_type=routing.get('answer_type') or 'symbolic')
elif source == 'GSM8K128':
routing = {'verifier_mode': 'numeric_equivalence', 'resolution': 'frozen_numeric'}
record.update(task_type='math', verifier_mode='numeric_equivalence', answer_type='numeric', allow_hash_answer=True)
solution = str(metadata.get('answer', '')).split('####', 1)[0].replace('\\n', '\n').strip()
unit = re.search('\\d\\s*(seconds?|minutes?|hours?)(?:/(?:day|week|month|year)s?)?[.\\s]*$', solution, re.I)
if unit:
record['reference_unit'] = unit.group(1).lower()
elif source in CHOICE_SOURCES:
normalized = choice_references(source, values)
record['reference_answer'] = normalized[0]
if len(normalized) > 1:
record['reference_answers'] = normalized
routing = {'verifier_mode': 'exact_choice', 'resolution': 'frozen_choice_with_source_index_normalization'}
record.update(task_type='choice', verifier_mode='exact_choice', answer_type='choice')
elif source == 'BBH128':
routing = {'verifier_mode': 'exact_text', 'resolution': 'frozen_exact_text'}
record.update(task_type='general', source_dataset='bbh', verifier_mode='exact_text', answer_type='exact_text')
elif source in {'IFEval128', 'IFBench128'}:
mode = 'ifeval' if source == 'IFEval128' else 'ifbench'
routing = {'verifier_mode': mode, 'resolution': 'official_instruction'}
record.update(task_type=mode, verifier_mode=mode, verifier_payload=metadata)
else:
raise ValueError(f'unsupported fixed evaluation source: {source}')
if source == 'BBH128':
record['accept_unmarked_final'] = accept_unmarked_final or '</think>' in response
if metadata.get('task') == 'word_sorting':
record['ordered_words'] = True
routing['resolution'] = 'bbh_ordered_words'
return (record, routing)
def code_payload(source: str, metadata: Mapping[str, Any], code: str) -> tuple[str, str]:
code = textwrap.dedent(code).strip()
if source == 'HumanEval':
entry = str(metadata['entry_point'])
if not re.search(f'\\bdef\\s+{re.escape(entry)}\\s*\\(', code):
code = str(metadata['prompt']).rstrip() + '\n' + textwrap.indent(code, ' ')
else:
context = re.split(f'(?m)^def\\s+{re.escape(entry)}\\s*\\(', str(metadata['prompt']), maxsplit=1)[0].strip()
if context:
code = context + '\n\n' + code
return (code, str(metadata['test']).rstrip() + f'\n\ncheck({entry})\n')
tests = [*metadata.get('test_list', []), *metadata.get('challenge_test_list', [])]
harness = '\n'.join((value for value in [str(metadata.get('test_setup_code') or ''), *map(str, tests)] if value.strip()))
return (code, harness)
@functools.lru_cache(maxsize=4)
def load_bbeh_evaluator(path: str):
with _BBEH_LOAD_LOCK:
spec = importlib.util.spec_from_file_location('edgerock_frozen_bbeh_evaluate', path)
if spec is None or spec.loader is None:
raise ImportError(f'cannot load BBEH evaluator: {path}')
module = importlib.util.module_from_spec(spec)
with contextlib.redirect_stdout(io.StringIO()):
spec.loader.exec_module(module)
return module.evaluate_correctness
def score_bbeh(row: Mapping[str, Any], visible: str, *, extract_final: bool=True) -> dict[str, Any]:
metadata = json.loads(str(row.get('metadata_json') or '{}'))
declared = typed_declared_answer(visible, answer_type='exact_text') if extract_final else None
ambiguous = bool(declared and declared['ambiguous'])
candidate = declared['candidate_answer'] if declared else visible.strip().splitlines()[-1] if extract_final and visible.strip() else visible
correct = bool(not ambiguous and load_bbeh_evaluator(str(OFFICIAL_RUNNERS / 'BBEH/evaluate.py'))(candidate, str(row['reference_answer'])))
return {'verdict': 'PASS' if correct else 'FAIL', 'method': 'bbeh_official_compare_declared_answer' if declared else 'bbeh_official_evaluate_correctness', 'candidate': declared, 'details': {'task': str(metadata['task']), 'revision': str(metadata['official_evaluator_revision'])}}
def execute_code(engine: VerifierEngine, row: Mapping[str, Any], metadata: Mapping[str, Any], code: str) -> dict[str, Any]:
source = str(row['source'])
if source == 'LiveCodeBenchV6':
candidate = textwrap.dedent(code).strip()
return engine.verify({'sample_id': f"{row['checkpoint_id']}:{row['source_sample_id']}:{row['sample_index']}", 'task_type': 'code', 'verifier_mode': 'code', 'problem': str(row['prompt']), 'response': f'</think>\n```python\n{candidate}\n```', 'finish_reason': 'eos', 'verifier_payload': {'language': 'python', 'lcb_payload_path': str(metadata['lcb_payload_path']), 'lcb_runner_path': str(OFFICIAL_RUNNERS / 'LiveCodeBench'), 'time_limit_s': 6}})
candidate, harness = code_payload(source, metadata, code)
return engine.verify({'sample_id': f"{row['checkpoint_id']}:{row['source_sample_id']}:{row['sample_index']}", 'task_type': 'code', 'verifier_mode': 'code', 'problem': str(row['prompt']), 'response': f'</think>\n```python\n{candidate}\n```', 'finish_reason': 'eos', 'verifier_payload': {'python_harness': harness, 'language': 'python', 'time_limit_s': 10}})
def score_code(engine: VerifierEngine, row: Mapping[str, Any], visible: str, *, accept_unfenced: bool=False) -> dict[str, Any]:
metadata = json.loads(str(row.get('metadata_json') or '{}'))
blocks = [value for value in code_blocks(visible) if str(value['language']).lower() in ACCEPTED_CODE_LANGUAGES]
strict_result = None
if blocks:
block = blocks[-1]
strict_result = execute_code(engine, row, metadata, visible[int(block['start']):int(block['end'])])
elif accept_unfenced and '```' not in visible:
strict_result = execute_code(engine, row, metadata, visible)
strict_pass = bool(strict_result and strict_result['verdict'] == 'PASS')
relaxed_result = strict_result
trimmed = TRAILING_IM_END.sub('', visible).rstrip()
opens = list(PYTHON_OPEN.finditer(trimmed))
repair_eligible = bool(row['finished_by_eos'] and (not row['cap_hit']) and (not blocks) and (len(opens) == 1) and (trimmed.count('```') == 1) and (opens[0].start() == 0))
if repair_eligible:
relaxed_result = execute_code(engine, row, metadata, trimmed[opens[0].end():])
return {'semantic_correct': bool(relaxed_result and relaxed_result['verdict'] == 'PASS'), 'strict_result': strict_result, 'relaxed_correct': bool(relaxed_result and relaxed_result['verdict'] == 'PASS'), 'relaxed_result': relaxed_result, 'repair_eligible': repair_eligible}
def instruction_counts(result: Mapping[str, Any]) -> tuple[int, int]:
external = (result.get('outcome') or {}).get('external') or {}
details = external.get('details') or {}
values = details.get('follow_instruction_list') or []
return (sum((bool(value) for value in values)), len(values))
def strict_gsm_correct(visible: str, reference: Any) -> bool:
candidate = last_balanced_boxed_answer(visible)
return bool(candidate is not None and compare_answers(candidate, [reference], answer_type='numeric_equivalence')['verdict'] == 'PASS')
def score_one(engine: VerifierEngine, row: Mapping[str, Any], think_open_id: int | None, think_close_id: int | None, *, prompt_think_closed: bool=False, allow_empty_reasoning: bool=False, accept_unmarked_final: bool=False) -> dict[str, Any]:
source_name = str(row['source'])
metadata = json.loads(str(row.get('metadata_json') or '{}'))
score_source = str(row.get('score_source') or metadata.get('score_source') or source_name)
accept_unmarked_final = accept_unmarked_final or bool(metadata.get('accept_unmarked_final'))
if source_name != score_source:
row = adapt_score_row({**row, 'score_source': score_source})
response = str(row['response'])
response_ids = row['response_ids']
finish_reason = evaluation_finish_reason(row)
think = think_protocol_status(response, finish_reason, response_ids, think_open_id=think_open_id, think_close_id=think_close_id, prompt_think_closed=prompt_think_closed, allow_empty_reasoning=allow_empty_reasoning)
mechanical = mechanical_repetitive_tail_metrics(response_ids, finish_reason)
source = str(row['source'])
instruction_pass = instruction_total = 0
relaxed_code = strict_code = None
code_status: Mapping[str, Any] | None = None
routing: Mapping[str, Any] = {}
if source in CODE_SOURCES or source == 'BBEH':
record = dict(response=response, problem=str(row['prompt']), task_type='code' if source in CODE_SOURCES else 'general', verifier_mode='code' if source in CODE_SOURCES else 'exact_text', finish_reason=finish_reason)
else:
record, routing = standard_verifier_record(row, response, accept_unmarked_final=accept_unmarked_final)
record['prompt_think_closed'] = prompt_think_closed
record = prepare_semantic_record(record, think_status=think, mechanical=mechanical['degenerate'])
visible = record_visible_response(record)
recovered_cot = record['answer_selection']['source'] == 'cot'
if source in CODE_SOURCES:
code_status = code_contract_status(response, response_ids, finish_reason, PYTHON_FENCE_CONTRACT, prompt_think_closed=prompt_think_closed)
if visible.strip():
code = score_code(engine, row, visible, accept_unfenced=accept_unmarked_final)
semantic = bool(code['semantic_correct'])
relaxed_code = bool(code['relaxed_correct'])
strict_code = bool(semantic and think['verdict'] == 'PASS' and code_status['contract_pass'] and (not mechanical['degenerate']) and (not recovered_cot))
verifier_result = code['relaxed_result'] or {'verdict': 'UNCERTAIN', 'method': 'no_complete_visible_python_fence'}
else:
semantic = relaxed_code = strict_code = False
verifier_result = missing_answer_outcome(record)
contract = strict_code
routing = {'verifier_mode': 'code', 'resolution': 'livecodebench_official' if source == 'LiveCodeBenchV6' else 'python_harness'}
elif source == 'BBEH':
verifier_result = score_bbeh(row, visible, extract_final=accept_unmarked_final or recovered_cot or prompt_think_closed or ('</think>' in response)) if visible.strip() else missing_answer_outcome(record)
semantic = verifier_result['verdict'] == 'PASS'
contract = bool(semantic and think['verdict'] == 'PASS' and (not mechanical['degenerate']) and (not recovered_cot))
routing = {'verifier_mode': 'bbeh_official', 'resolution': 'google_deepmind_evaluate_correctness'}
else:
verifier_result = engine.verify(record)
semantic = verifier_result['verdict'] == 'PASS'
contract = bool(semantic and think['verdict'] == 'PASS' and (not mechanical['degenerate']) and (not recovered_cot))
if source in {'IFEval128', 'IFBench128'}:
instruction_pass, instruction_total = instruction_counts(verifier_result)
verifier_result.setdefault('outcome', {})['answer_selection'] = record['answer_selection']
if recovered_cot:
verifier_result['outcome']['format_kind'] = 'missing_final'
original_label_correct = None
primary_view_eligible = None
if source == 'MMLURedux':
primary_view_eligible = bool(metadata['primary_corrected_choice_eligible'])
original_record = dict(record)
original_record['reference_answer'] = str(metadata['original_answer_letter'])
original_record.pop('reference_answers', None)
original_label_correct = bool(visible.strip() and engine.verify(original_record)['verdict'] == 'PASS')
subgroup = ''
if source in MACRO_AVERAGE_SUBSET_FIELDS:
subgroup = str(metadata[MACRO_AVERAGE_SUBSET_FIELDS[source]])
elif source == 'MMLUPro':
subgroup = str(metadata['category'])
elif source == 'MMLURedux':
subgroup = str(metadata['annotation_error_type'])
elif source == 'SuperGPQA':
subgroup = str(metadata['discipline'])
elif source == 'LiveCodeBenchV6':
subgroup = str(metadata['difficulty'])
elif source == 'HMMT2026':
subgroup = ','.join(map(str, metadata['problem_type']))
gsm = strict_gsm_correct(visible, row['reference_answer']) if source == 'GSM8K128' else None
return {'checkpoint_id': str(row['checkpoint_id']), 'protocol_key': str(row['protocol_key']), 'source': source_name, 'report_source': source_name, 'benchmark_subgroup': subgroup, 'source_sample_id': str(row['source_sample_id']), 'problem_family_id': str(row['problem_family_id']), 'sample_index': int(row['sample_index']), 'semantic_correct': bool(semantic), 'original_label_correct': original_label_correct, 'primary_view_eligible': primary_view_eligible, 'contract_correct': bool(contract), 'strict_code_correct': strict_code, 'relaxed_code_correct': relaxed_code, 'strict_gsm_correct': gsm, 'instruction_pass': instruction_pass, 'instruction_total': instruction_total, 'verdict': str(verifier_result['verdict']), 'method': str(verifier_result.get('method') or ''), 'effective_verifier_mode': str(routing.get('verifier_mode') or ''), 'routing_resolution': str(routing.get('resolution') or ''), 'routing_reason': str(routing.get('reason') or ''), 'think_verdict': str(think['verdict']), 'think_reason': str(think['reason']), 'think_open_count': int(think['opening_count']), 'think_close_count': int(think['closing_count']), 'answer_visible': bool(visible) and (not recovered_cot), 'finished_by_eos': bool(row['finished_by_eos']), 'cap_hit': bool(row['cap_hit']), 'mechanical_repetition': bool(mechanical['degenerate']), 'mechanical_onset_index': mechanical['failure_onset_index'], 'code_contract_reason': '' if code_status is None else str(code_status['contract_reason']), 'generated_tokens': int(row['generated_tokens']), 'scoring_json': json.dumps(verifier_result, ensure_ascii=False, sort_keys=True)}
def build_engine(config: Mapping[str, Any], *, external_factory=ExternalVerifier) -> VerifierEngine:
project = Path(__file__).resolve().parents[1]
verifier = config['execution_contract']['verifier_service']
assets = Path(verifier['external_verifier_assets'])
python = shlex.quote(sys.executable)
math_command = f"{python} {shlex.quote(str(project.parent / 'run.py'))}"
code_command = f"{python} {shlex.quote(str(project.parent / 'run.py'))}"
instruction_command = ' '.join(map(shlex.quote, [sys.executable, str(project.parent / 'run.py'), '_verify', 'instruction', '--ifbench-repo', str(assets / 'ifbench_official'), '--ifeval-repo', str(assets / 'ifeval_official'), '--nltk-data', str(assets / 'nltk_data')]))
math_command += ' _verify math'
code_command += ' _verify code'
timeout = float(verifier['external_timeout_seconds'])
return VerifierEngine(math_verifier=external_factory(math_command, timeout=timeout, worker_count=int(verifier['math_workers'])), code_verifier=external_factory(code_command, timeout=None, worker_count=int(verifier['code_workers'])), instruction_verifier=external_factory(instruction_command, timeout=timeout, worker_count=int(verifier['instruction_workers'])), service_instance_id='fixed-eval-frozen-protocol')
class _ExternalVerifierProxy(BaseProxy):
_exposed_ = ('run', 'close', '__getattribute__')
@property
def enabled(self):
return self._callmethod('__getattribute__', ('enabled',))
@property
def worker_count(self):
return self._callmethod('__getattribute__', ('worker_count',))
def run(self, request):
return self._callmethod('run', (request,))
def close(self):
return self._callmethod('close')
class _ScoringVerifierManager(BaseManager):
pass
@contextlib.contextmanager
def shared_scoring_engine(config, context):
with _ScoringVerifierManager(ctx=context) as manager:
engine = build_engine(config, external_factory=manager.ExternalVerifier)
try:
yield engine
finally:
engine.close()
def percentile(values: list[int], value: float) -> float:
return float(np.percentile(np.asarray(values, dtype=np.float64), value))
def pass_estimate(correct: int, samples: int, k: int) -> float:
if correct <= 0:
return 0.0
if samples - correct < k:
return 1.0
return 1.0 - math.comb(samples - correct, k) / math.comb(samples, k)
def summarize(records: list[Mapping[str, Any]]) -> dict[str, Any]:
source = str(records[0]['report_source'])
cases: dict[str, list[Mapping[str, Any]]] = defaultdict(list)
for record in records:
cases[str(record['source_sample_id'])].append(record)
tokens = [int(record['generated_tokens']) for record in records]
attempts = len(records)
semantic = sum((bool(record['semantic_correct']) for record in records))
contract = sum((bool(record['contract_correct']) for record in records))
sample_counts = sorted({len(value) for value in cases.values()})
relaxed = [bool(record['relaxed_code_correct']) for record in records if record['relaxed_code_correct'] is not None]
strict_code = [bool(record['strict_code_correct']) for record in records if record['strict_code_correct'] is not None]
strict_gsm = [bool(record['strict_gsm_correct']) for record in records if record['strict_gsm_correct'] is not None]
instruction_total = sum((int(record['instruction_total']) for record in records))
subgroups: dict[str, list[Mapping[str, Any]]] = defaultdict(list)
for record in records:
if str(record.get('benchmark_subgroup') or ''):
subgroups[str(record['benchmark_subgroup'])].append(record)
subgroup_semantic = {name: statistics.fmean((bool(record['semantic_correct']) for record in values)) for name, values in sorted(subgroups.items())}
primary_records = [record for record in records if record.get('primary_view_eligible') is True]
original_label = [bool(record['original_label_correct']) for record in records if record.get('original_label_correct') is not None]
clean_redux = [bool(record['semantic_correct']) for record in records if record.get('benchmark_subgroup') == 'ok']
observed = {metric: statistics.fmean((any((bool(row[metric]) for row in value)) for value in cases.values())) for metric in ('semantic_correct', 'contract_correct')}
pass_ks = [k for k in sorted({1, 2, 4, 8, *sample_counts}) if all((k <= len(value) for value in cases.values()))]
execution_counts = Counter()
for row in records:
outcome = json.loads(row.get('scoring_json') or '{}').get('outcome') or {}
details = (outcome.get('external') or {}).get('details') or {}
if isinstance(details, dict):
for name in ('global_timeout', 'missing_result', 'missing_metadata', 'scorer_error'):
if details.get(name) is True:
execution_counts[name] += 1
result = {'attempts': attempts, 'upstream_execution_counts': dict(execution_counts), 'cases': len(cases), 'samples_per_case': sample_counts, 'semantic_mean_pass_at_1': semantic / attempts, 'contract_mean_pass_at_1': contract / attempts, 'semantic_observed_pass_at_k': observed['semantic_correct'], 'contract_observed_pass_at_k': observed['contract_correct'], 'unbiased_pass_at_k': {metric: {str(k): statistics.fmean((pass_estimate(sum((bool(row[metric]) for row in value)), len(value), k) for value in cases.values())) for k in pass_ks} for metric in ('semantic_correct', 'contract_correct')}, 'strict_code_mean_pass_at_1': statistics.fmean(strict_code) if strict_code else None, 'relaxed_code_mean_pass_at_1': statistics.fmean(relaxed) if relaxed else None, 'strict_gsm_accuracy': statistics.fmean(strict_gsm) if strict_gsm else None, 'instruction_level_accuracy': sum((int(record['instruction_pass']) for record in records)) / instruction_total if instruction_total else None, 'all_correct_cases': sum((all((bool(row['semantic_correct']) for row in value)) for value in cases.values())), 'mixed_cases': sum((any((bool(row['semantic_correct']) for row in value)) and (not all((bool(row['semantic_correct']) for row in value))) for value in cases.values())), 'all_wrong_cases': sum((not any((bool(row['semantic_correct']) for row in value)) for value in cases.values())), 'thinking_closed_exactly_once_rate': statistics.fmean((int(record['think_close_count'] == 1) for record in records)), 'thinking_protocol_pass_rate': statistics.fmean((record['think_verdict'] == 'PASS' for record in records)), 'thinking_verdicts': dict(Counter((record['think_verdict'] for record in records))), 'thinking_reasons': dict(Counter((record['think_reason'] for record in records))), 'eos_rate': statistics.fmean((bool(record['finished_by_eos']) for record in records)), 'cap_hit_rate': statistics.fmean((bool(record['cap_hit']) for record in records)), 'answer_visible_rate': statistics.fmean((bool(record['answer_visible']) for record in records)), 'mechanical_repetition_rate': statistics.fmean((bool(record['mechanical_repetition']) for record in records)), 'mean_generated_tokens': statistics.fmean(tokens), 'p50_generated_tokens': percentile(tokens, 50), 'p90_generated_tokens': percentile(tokens, 90), 'p95_generated_tokens': percentile(tokens, 95), 'p99_generated_tokens': percentile(tokens, 99), 'maximum_generated_tokens': max(tokens), 'semantic_correct_per_million_tokens': semantic * 1000000 / sum(tokens) if sum(tokens) else None, 'verdicts': dict(Counter((str(record['verdict']) for record in records))), 'methods': dict(Counter((str(record['method']) for record in records))), 'effective_verifier_modes': dict(Counter((str(record['effective_verifier_mode']) for record in records))), 'routing_resolutions': dict(Counter((str(record['routing_resolution']) for record in records)))}
result['subgroup_semantic_mean_pass_at_1'] = subgroup_semantic
result['subgroup_semantic_macro_pass_at_1'] = statistics.fmean(subgroup_semantic.values()) if subgroup_semantic else None
if source in MACRO_AVERAGE_SUBSET_FIELDS:
result['official_primary_mean_pass_at_1'] = result['subgroup_semantic_macro_pass_at_1']
elif source == 'MMLURedux':
result['official_primary_mean_pass_at_1'] = statistics.fmean((bool(record['semantic_correct']) for record in primary_records)) if primary_records else None
result['original_label_accuracy'] = statistics.fmean(original_label) if original_label else None
result['clean_accuracy'] = statistics.fmean(clean_redux) if clean_redux else None
result['corrected_choice_accuracy'] = result['official_primary_mean_pass_at_1']
result['corrected_choice_attempts'] = len(primary_records)
else:
result['official_primary_mean_pass_at_1'] = result['semantic_mean_pass_at_1']
return result
def last_balanced_boxed_answer(response_text: str) -> str | None:
text = str(response_text)
values: list[str] = []
for match in re.finditer('\\\\boxed\\s*\\{', text):
start = match.end()
depth = 1
for index in range(start, len(text)):
character = text[index]
if character == '{':
depth += 1
elif character == '}':
depth -= 1
if depth == 0:
value = text[start:index].strip()
if value:
values.append(value)
break
return values[-1] if values else None
_ScoringVerifierManager.register('ExternalVerifier', ExternalVerifier, _ExternalVerifierProxy)
|