File size: 19,774 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 | """Frozen, calibration-only selection and paired analysis for Experiment 001B.
This module has no inference dependencies. It expects the per-generation JSON
records saved by the 001B runner and makes incomplete calibration coverage
visible instead of ranking the surviving candidates.
"""
from __future__ import annotations
from collections.abc import Iterable, Mapping, Sequence
from typing import Any
from paired_stats import paired_cluster_summary
_NATURAL_FINISHES = {"stop", "eos", "end_of_sequence", "natural_stop"}
_SCORABLE_STATUSES = {"correct", "incorrect"}
_KNOWN_SCORE_STATUSES = _SCORABLE_STATUSES | {"unresolved", "unparsed", "error"}
_DEFAULT_FORBIDDEN_CANDIDATES = {"base", "base_thinking", "base_no_thinking"}
DEFAULT_CONDITIONS = (
"base",
"edited",
"published_constant",
"independent_constant",
"norm_matched_random",
"rank1_activation_dependent",
)
def _record_key(record: Mapping[str, Any]) -> tuple[int, int]:
return int(record["source_row"]), int(record["seed"])
def _score_status(record: Mapping[str, Any]) -> str | None:
score = record.get("score")
if not isinstance(score, Mapping):
return None
value = score.get("status")
return str(value) if value is not None else None
def _score_issue(record: Mapping[str, Any]) -> str | None:
"""Return a data/scoring error instead of silently awarding model no-credit.
The frozen grader emits one of: correct, incorrect, unresolved, unparsed,
or error, with a matching tri-state ``correct`` value. Missing or
inconsistent fields are malformed score records.
"""
score = record.get("score")
if not isinstance(score, Mapping):
return "missing_score_object"
status = score.get("status")
if not isinstance(status, str) or status not in _KNOWN_SCORE_STATUSES:
return "unknown_or_missing_score_status"
correct = score.get("correct", _MISSING)
if correct is _MISSING:
return "missing_correct_flag"
if status == "correct" and correct is not True:
return "correct_status_mismatch"
if status == "incorrect" and correct is not False:
return "incorrect_status_mismatch"
if status in {"unresolved", "unparsed", "error"} and correct is not None:
return "unscorable_status_mismatch"
return None
_MISSING = object()
def _record_problem(record: Mapping[str, Any]) -> str | None:
if record.get("error"):
return "technical_error"
issue = _score_issue(record)
if issue is not None:
return f"scoring_error:{issue}"
if _score_status(record) == "error":
score = record.get("score")
detail = score.get("error") if isinstance(score, Mapping) else None
return f"scoring_error:{detail or 'grader_error_status'}"
return None
def record_validation_issue(record: Mapping[str, Any]) -> str | None:
"""Public record-integrity check used by analysis summaries and CSV output."""
return _record_problem(record)
def _is_natural_completion(record: Mapping[str, Any]) -> bool:
if record.get("truncated") is True:
return False
reason = str(record.get("finish_reason", "")).lower()
if reason in _NATURAL_FINISHES:
return True
return bool(record.get("stopped_on_eos", False))
def _capped_length(record: Mapping[str, Any] | None, max_new_tokens: int) -> int:
if record is None:
return max_new_tokens
value = record.get("tokens", record.get("token_count"))
if value is None:
return max_new_tokens
return min(max_new_tokens, max(0, int(value)))
def select_calibration_winners(
records_by_candidate: Mapping[str, Iterable[Mapping[str, Any]]],
families: Mapping[str, Sequence[str]],
expected_question_seeds: Sequence[tuple[int, int]],
*,
max_new_tokens: int,
forbidden_candidates: Iterable[str] = _DEFAULT_FORBIDDEN_CANDIDATES,
) -> dict[str, Any]:
"""Select one fully observed candidate per family using calibration only.
Ranking, from best to worst: maximize verified solves; maximize natural
completions; minimize unscorable outputs; minimize mean capped generated
length; lexical candidate ID. Any missing record, technical error, or
malformed grade blocks that whole family (no fallback winner); callers
should resume the run or report the family incomplete. A base comparator
cannot be a candidate.
"""
if max_new_tokens <= 0:
raise ValueError("max_new_tokens must be positive")
if not expected_question_seeds:
raise ValueError("expected_question_seeds cannot be empty")
expected = [(int(row), int(seed)) for row, seed in expected_question_seeds]
if len(expected) != len(set(expected)):
raise ValueError("expected question/seed keys must be unique")
forbidden = set(forbidden_candidates)
indexed: dict[str, dict[tuple[int, int], Mapping[str, Any]]] = {}
for candidate, records in records_by_candidate.items():
if candidate in forbidden or candidate.lower() in forbidden:
raise ValueError(f"Comparator {candidate!r} cannot be a calibration candidate")
by_key: dict[tuple[int, int], Mapping[str, Any]] = {}
for record in records:
key = _record_key(record)
if key in by_key:
raise ValueError(f"Duplicate calibration record for {candidate}/{key}")
if record.get("candidate_id", candidate) != candidate:
raise ValueError(f"Record candidate_id disagrees with mapping key: {candidate}/{key}")
by_key[key] = record
extra = set(by_key) - set(expected)
if extra:
raise ValueError(f"Unexpected calibration keys for {candidate}: {sorted(extra)}")
indexed[candidate] = by_key
all_family_candidates = {candidate for values in families.values() for candidate in values}
if not all_family_candidates:
raise ValueError("families must contain at least one candidate")
if sum(len(values) for values in families.values()) != len(all_family_candidates):
raise ValueError("a candidate ID cannot appear in multiple calibration families")
if all_family_candidates & forbidden:
raise ValueError("Base/comparator cannot be selected in a family")
missing_candidate_records = all_family_candidates - set(indexed)
family_results: dict[str, Any] = {}
for family, candidates_seq in families.items():
candidates = list(candidates_seq)
if not candidates or len(candidates) != len(set(candidates)):
raise ValueError(f"Family {family!r} must list unique candidates")
candidate_metrics: dict[str, Any] = {}
family_blockers: list[str] = []
for candidate in candidates:
by_key = indexed.get(candidate, {})
missing = sorted(set(expected) - set(by_key))
error_keys = sorted(
key for key, record in by_key.items()
if _record_problem(record) == "technical_error"
)
score_error_keys = sorted(
key for key, record in by_key.items()
if _record_problem(record) is not None
and _record_problem(record) != "technical_error"
)
if candidate in missing_candidate_records:
family_blockers.append(f"missing_candidate:{candidate}")
if missing:
family_blockers.append(f"missing_records:{candidate}")
if error_keys:
family_blockers.append(f"technical_errors:{candidate}")
if score_error_keys:
family_blockers.append(f"scoring_errors:{candidate}")
solves = sum(
key in by_key
and _record_problem(by_key[key]) is None
and _score_status(by_key[key]) == "correct"
for key in expected
)
natural = sum(_is_natural_completion(by_key[key]) for key in expected if key in by_key)
unscorable = sum(
_score_status(by_key.get(key, {})) not in _SCORABLE_STATUSES
for key in expected
)
lengths = [_capped_length(by_key.get(key), max_new_tokens) for key in expected]
candidate_metrics[candidate] = {
"n_expected": len(expected),
"n_recorded": len(by_key),
"n_missing": len(missing),
"missing_question_seeds": [list(key) for key in missing],
"n_technical_errors": len(error_keys),
"technical_error_question_seeds": [list(key) for key in error_keys],
"n_scoring_errors": len(score_error_keys),
"scoring_error_question_seeds": [list(key) for key in score_error_keys],
"solves": solves,
"natural_completions": natural,
"unscorable": unscorable,
"mean_capped_tokens": sum(lengths) / len(expected),
"complete": not missing and not error_keys and not score_error_keys and candidate in indexed,
}
complete = not family_blockers
winner = None
if complete:
winner = min(
candidates,
key=lambda candidate: (
-candidate_metrics[candidate]["solves"],
-candidate_metrics[candidate]["natural_completions"],
candidate_metrics[candidate]["unscorable"],
candidate_metrics[candidate]["mean_capped_tokens"],
candidate,
),
)
family_results[family] = {
"status": "selected" if complete else "incomplete",
"selected_candidate_id": winner,
"blockers": sorted(set(family_blockers)),
"candidates": candidate_metrics,
"selection_rule": [
"maximize solves",
"maximize natural completions",
"minimize unscorable outputs",
"minimize mean capped generated tokens",
"lexical candidate ID",
],
"selection_source": "calibration only",
}
return {
"phase": "calibration",
"max_new_tokens": max_new_tokens,
"expected_question_seeds": [list(key) for key in expected],
"families": family_results,
"base_is_comparator_only": True,
"heldout_used_for_selection": False,
}
def collect_paired_outcomes(
records: Iterable[Mapping[str, Any]],
condition: str,
candidate_id: str | None = None,
) -> tuple[dict[int, dict[int, bool]], dict[tuple[int, int], str]]:
"""Index one condition's valid records; return technical/scoring exclusions too."""
outcomes: dict[int, dict[int, bool]] = {}
excluded: dict[tuple[int, int], str] = {}
seen: set[tuple[int, int]] = set()
for record in records:
if record.get("condition") != condition:
continue
if candidate_id is not None and record.get("candidate_id") != candidate_id:
continue
key = _record_key(record)
if key in seen:
raise ValueError(f"Duplicate heldout record for {condition}/{key}")
seen.add(key)
question, seed = key
outcomes.setdefault(question, {})
problem = _record_problem(record)
if problem is not None:
excluded[key] = problem
continue
status = _score_status(record)
# For author-aligned solve-by-cap accuracy, non-completion and
# unscorable model outputs are no-credit outcomes, not technical errors.
outcomes[question][seed] = status == "correct" or (
isinstance(record.get("score"), Mapping)
and record["score"].get("correct") is True
)
return outcomes, excluded
def paired_heldout_summary(
records: Iterable[Mapping[str, Any]],
first_condition: str,
comparator_condition: str,
expected_question_seeds: Sequence[tuple[int, int]],
*,
first_candidate_id: str | None = None,
comparator_candidate_id: str | None = None,
bootstrap_seed: int = 14353,
bootstrap_samples: int = 10_000,
) -> dict[str, Any]:
"""Compare two heldout conditions and expose every missing/error pair."""
if first_condition == comparator_condition:
raise ValueError("first_condition and comparator_condition must differ")
expected_list = [(int(question), int(seed)) for question, seed in expected_question_seeds]
if len(expected_list) != len(set(expected_list)):
raise ValueError("expected_question_seeds must be unique")
rows = list(records)
first, first_errors = collect_paired_outcomes(rows, first_condition, first_candidate_id)
comparator, comparator_errors = collect_paired_outcomes(
rows, comparator_condition, comparator_candidate_id
)
expected = set(expected_list)
first_keys = {(question, seed) for question, seeds in first.items() for seed in seeds}
comparator_keys = {(question, seed) for question, seeds in comparator.items() for seed in seeds}
first_present = first_keys | set(first_errors)
comparator_present = comparator_keys | set(comparator_errors)
unexpected = (first_present | comparator_present) - expected
if unexpected:
raise ValueError(f"Unexpected heldout question/seed keys: {sorted(unexpected)}")
matched_keys = sorted((first_keys & comparator_keys) & expected)
first_map: dict[int, dict[int, bool]] = {}
comparator_map: dict[int, dict[int, bool]] = {}
for question, seed in matched_keys:
first_map.setdefault(question, {})[seed] = first[question][seed]
comparator_map.setdefault(question, {})[seed] = comparator[question][seed]
summary = paired_cluster_summary(
first_map,
comparator_map,
bootstrap_seed=bootstrap_seed,
bootstrap_samples=bootstrap_samples,
)
incomplete = []
for key in sorted(expected):
missing = []
if key not in first_present:
missing.append(first_condition)
if key not in comparator_present:
missing.append(comparator_condition)
reasons = {
first_condition: first_errors.get(key),
comparator_condition: comparator_errors.get(key),
}
if missing or any(reasons.values()):
technical = {
name: value for name, value in reasons.items() if value == "technical_error"
}
item = {
"source_row": key[0],
"seed": key[1],
"missing_conditions": missing,
"technical_errors": technical,
}
scoring = {
name: value for name, value in reasons.items()
if value and value != "technical_error"
}
if technical:
item["technical_errors"] = technical
if scoring:
item["scoring_errors"] = scoring
incomplete.append(item)
summary.update(
{
"first_condition": first_condition,
"comparator_condition": comparator_condition,
"first_candidate_id": first_candidate_id,
"comparator_candidate_id": comparator_candidate_id,
"expected_generation_pairs": len(expected),
"complete_matched_generation_pairs": len(matched_keys),
"incomplete_pairs": incomplete,
"analysis_uses_only_complete_pairs": True,
"endpoint": "verified correct final answer; unresolved and unparsed are no-credit",
}
)
if len(matched_keys) != len(expected):
summary["question_bootstrap_percentile_95"] = None
summary["exact_mcnemar_two_sided_p"] = None
summary["inference_suppressed"] = (
"Missing, technical-error, or scoring-error pairs make complete-case inferential summaries unsafe; report the matched counts descriptively."
)
return summary
def summarize_experiment_001b_contrasts(
records: Iterable[Mapping[str, Any]],
expected_question_seeds: Sequence[tuple[int, int]],
condition_config_ids: Mapping[str, str] | None = None,
*,
bootstrap_seed: int = 14353,
bootstrap_samples: int = 10_000,
) -> dict[str, Any]:
"""Build the frozen primary and explicitly exploratory 001B contrasts.
The single prespecified contrast is edited checkpoint vs the
calibration-selected published-constant direction. Remaining comparisons
are diagnostics or intervention-vs-base summaries and their p-values are
exploratory and unadjusted for multiplicity.
"""
ids = dict(condition_config_ids or {})
rows = list(records)
def summarize(first: str, comparator: str) -> dict[str, Any]:
return paired_heldout_summary(
rows,
first,
comparator,
expected_question_seeds,
first_candidate_id=ids.get(first, first),
comparator_candidate_id=ids.get(comparator, comparator),
bootstrap_seed=bootstrap_seed,
bootstrap_samples=bootstrap_samples,
)
base_conditions = [
"edited",
"published_constant",
"independent_constant",
"norm_matched_random",
"rank1_activation_dependent",
]
diagnostics = {
"rank1_activation_dependent_minus_published_constant": (
"rank1_activation_dependent",
"published_constant",
),
"rank1_activation_dependent_minus_norm_matched_random": (
"rank1_activation_dependent",
"norm_matched_random",
),
"edited_minus_rank1_activation_dependent": (
"edited",
"rank1_activation_dependent",
),
}
return {
"plan": {
"primary": "edited_minus_published_constant",
"primary_definition": (
"edited checkpoint vs the published_constant candidate selected using calibration only"
),
"other_comparisons": "exploratory; p-values are unadjusted and not confirmatory",
"small_heldout_note": (
"Ten heldout questions support a mechanism screen only; do not claim population accuracy effects or equivalence."
),
},
"primary_prespecified": {
"comparison_id": "edited_minus_published_constant",
"first_condition": "edited",
"comparator_condition": "published_constant",
"summary": summarize("edited", "published_constant"),
},
"exploratory_vs_base_unadjusted": {
f"{condition}_minus_base": summarize(condition, "base")
for condition in base_conditions
},
"exploratory_diagnostic_pairs_unadjusted": {
comparison_id: {
"first_condition": pair[0],
"comparator_condition": pair[1],
"summary": summarize(*pair),
}
for comparison_id, pair in diagnostics.items()
},
}
__all__ = [
"DEFAULT_CONDITIONS",
"collect_paired_outcomes",
"paired_heldout_summary",
"record_validation_issue",
"select_calibration_winners",
"summarize_experiment_001b_contrasts",
]
|