Datasets:
File size: 17,334 Bytes
e1ced61 | 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 | """Common normalized item schema, content-addressed images, and the eval registry.
Stage 1 of the data pipeline (``docs/02_DATA_PIPELINE.md`` §4) maps every source
row to one immutable :class:`NormalizedItem` matching
``schemas/normalized_item.schema.json``. The canonical ``base_id`` is the
content-addressed hash of ``docs/02`` §3.1 (source identity + question +
choices + image family); ``question_sha256`` / ``choices_sha256`` are over the
exact UTF-8 source bytes (the model input string is never normalized). Images
are content-addressed by their SHA-256 so the same bytes are stored once and
referenced by hash from any mount.
Stage 0 (``docs/02`` §3) freezes the evaluation registry
(``evaluation_items.v1.jsonl``) *before* any training source is touched. The
registry is write-once: a second freeze that would change its bytes is a hard
failure, never a silent overwrite.
"""
from __future__ import annotations
from collections.abc import Callable, Iterable, Iterator, Mapping, Sequence
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Literal
from ..atomic_io import JsonlAppender, atomic_write_bytes, atomic_write_jsonl, read_jsonl
from ..hashing import (
base_id as compute_base_id,
)
from ..hashing import (
canonical_json,
choices_sha256,
sha256_bytes,
)
from ..hashing import (
image_family_sha256 as compute_image_family_sha256,
)
from ..hashing import question_sha256 as compute_question_sha256
from ..paths import repo_root
from ..schema_io import load_schema, validate
SCHEMA_VERSION = 2
AnswerType = Literal["multiple_choice", "integer", "number", "expression", "short_text", "boolean"]
# ADR-0002 normalized-item policies (schemas/normalized_item.schema.json v2).
# C1 rows come from source-native structured worlds (PlotQA, Geometry3K); C2 rows
# are dual-compiled train candidates; untouched rows are retention-eval only;
# blocked rows are held back and never substituted.
Policy = Literal[
"c1_train_candidate",
"c1_certified_eval_candidate",
"c2_train_candidate",
"untouched_evaluation_only",
"blocked",
]
# A resolver turns one native image reference (a path/relative name in the raw
# row) into its raw bytes. The importer never opens files itself, so importers
# are unit-testable with an in-memory resolver.
ImageResolver = Callable[[str], bytes]
_NORMALIZED_ITEM_SCHEMA: dict[str, Any] | None = None
class IngestError(ValueError):
"""Raised when a source row violates an ingest invariant.
Invariants are hard failures (never silently dropped): a non-math MMK12 row
is source revision drift, not a row to skip.
"""
class RegistryFrozenError(IngestError):
"""Raised when a write-once registry would be changed by a re-freeze."""
def normalized_item_schema() -> dict[str, Any]:
"""Load and cache the compiled ``normalized_item`` JSON Schema."""
global _NORMALIZED_ITEM_SCHEMA
if _NORMALIZED_ITEM_SCHEMA is None:
_NORMALIZED_ITEM_SCHEMA = load_schema(
repo_root() / "schemas" / "normalized_item.schema.json"
)
return _NORMALIZED_ITEM_SCHEMA
def validate_normalized_item(row: Mapping[str, Any]) -> None:
"""Raise :class:`IngestError` if ``row`` violates the normalized item schema."""
errors = validate(dict(row), normalized_item_schema())
if errors:
raise IngestError("normalized item failed schema validation: " + "; ".join(errors))
# --- content-addressed image store -----------------------------------------
def _ext_for(data: bytes) -> str:
"""Infer a file extension from image bytes (PNG/JPEG), else ``bin``."""
if data.startswith(b"\x89PNG\r\n\x1a\n"):
return "png"
if data.startswith(b"\xff\xd8\xff"):
return "jpg"
return "bin"
@dataclass
class ImageStore:
"""Store image bytes content-addressed under ``<output>/images/<sha>.<ext>``.
The same bytes are written once; repeat stores are a no-op. Paths are
returned relative to ``base_dir`` so artifacts reproduce across mounts.
"""
base_dir: Path
def __post_init__(self) -> None:
(self.base_dir / "images").mkdir(parents=True, exist_ok=True)
def store(self, data: bytes) -> tuple[str, str]:
"""Store ``data``; return ``(relative_path, sha256)``."""
digest = sha256_bytes(data)
ext = _ext_for(data)
relative = f"images/{digest}.{ext}"
target = self.base_dir / relative
if not target.exists():
atomic_write_bytes(target, data)
return relative, digest
class DirectoryImageResolver:
"""Resolve native image refs to bytes by reading from a root directory."""
def __init__(self, root: str | Path) -> None:
self.root = Path(root)
def __call__(self, ref: str) -> bytes:
return (self.root / ref).read_bytes()
# --- normalized item -------------------------------------------------------
@dataclass(frozen=True)
class Choice:
"""One multiple-choice option in source order."""
key: str
text: str
def to_dict(self) -> dict[str, str]:
return {"key": self.key, "text": self.text}
@dataclass(frozen=True)
class NormalizedItem:
"""One immutable normalized source row (``docs/02`` §4 schema)."""
base_id: str
source: str
source_revision: str
source_config: str
source_split: str
source_native_id: str
question: str
question_sha256: str
choices: tuple[Choice, ...]
choices_sha256: str
answer_raw: str | int | bool
answer_canonical: str | int | bool
answer_type: AnswerType
image_paths: tuple[str, ...]
image_sha256: tuple[str, ...]
policy: Policy
provenance: dict[str, Any]
subject: str | None = None
license_gate: str | None = None
def to_row(self) -> dict[str, Any]:
"""Serialize to the schema-conforming dict written to ``items.jsonl``."""
row: dict[str, Any] = {
"schema_version": SCHEMA_VERSION,
"base_id": self.base_id,
"source": self.source,
"source_revision": self.source_revision,
"source_config": self.source_config,
"source_split": self.source_split,
"source_native_id": self.source_native_id,
"question": self.question,
"question_sha256": self.question_sha256,
"choices": [c.to_dict() for c in self.choices],
"choices_sha256": self.choices_sha256,
"answer_raw": self.answer_raw,
"answer_canonical": self.answer_canonical,
"answer_type": self.answer_type,
"image_paths": list(self.image_paths),
"image_sha256": list(self.image_sha256),
"provenance": dict(self.provenance),
"policy": self.policy,
}
if self.subject is not None:
row["subject"] = self.subject
if self.license_gate is not None:
row["license_gate"] = self.license_gate
return row
def _native_row_sha256(row: Mapping[str, Any]) -> str:
"""SHA-256 of the native row's canonical JSON (the source-record fingerprint)."""
return sha256_bytes(canonical_json(dict(row)).encode("utf-8"))
def canonicalize_mc_answer(answer: str, choices: Sequence[Choice]) -> str | int | bool:
"""Map a multiple-choice answer to its choice key when it matches a text.
Sources disagree on whether the answer is a key (``"B"``) or the option
text (``"42"``). The canonical form is the choice key when the raw answer
equals one option's text; otherwise the raw string is preserved.
"""
text_answer = str(answer).strip()
for choice in choices:
if text_answer == str(choice.text).strip():
return choice.key
return text_answer
def make_item(
*,
source: str,
source_revision: str,
source_config: str,
source_split: str,
source_native_id: str,
question: str,
choices: Sequence[Choice] | Sequence[Mapping[str, str]],
answer_raw: str | int | bool,
answer_canonical: str | int | bool,
answer_type: AnswerType,
image_paths: Sequence[str],
image_sha256: Sequence[str],
policy: Policy,
native_row: Mapping[str, Any],
subject: str | None = None,
license_gate: str | None = None,
extra_provenance: Mapping[str, Any] | None = None,
) -> NormalizedItem:
"""Build a :class:`NormalizedItem`, computing ``base_id`` and byte hashes.
``base_id`` is the content-addressed hash of ``docs/02`` §3.1 —
sha256(canonical_json({source, source_revision, source_native_id,
image_family_sha256, question_sha256, choices_sha256})). ``source_config``
and ``source_split`` are stored on the item but excluded from the ID. The
native row's canonical-JSON SHA-256 is recorded in provenance so a row can
be traced back to its source record.
"""
if not question:
raise IngestError(f"{source}: question must be non-empty")
if not image_paths or len(image_paths) != len(image_sha256):
raise IngestError(
f"{source}/{source_native_id}: image_paths and image_sha256 must be "
"parallel non-empty sequences"
)
norm_choices = tuple(
c if isinstance(c, Choice) else Choice(key=str(c["key"]), text=str(c["text"]))
for c in choices
)
q_sha = compute_question_sha256(question)
c_sha = choices_sha256([{"key": c.key, "text": c.text} for c in norm_choices])
fam_sha = compute_image_family_sha256(image_sha256)
bid = compute_base_id(
source=source,
source_revision=source_revision,
source_native_id=source_native_id,
image_family_sha256=fam_sha,
question_sha256=q_sha,
choices_sha256=c_sha,
)
provenance: dict[str, Any] = {
"native_row_json_sha256": _native_row_sha256(native_row),
}
if extra_provenance:
provenance.update(extra_provenance)
item = NormalizedItem(
base_id=bid,
source=source,
source_revision=source_revision,
source_config=source_config,
source_split=source_split,
source_native_id=str(source_native_id),
question=question,
question_sha256=q_sha,
choices=norm_choices,
choices_sha256=c_sha,
answer_raw=answer_raw,
answer_canonical=answer_canonical,
answer_type=answer_type,
image_paths=tuple(image_paths),
image_sha256=tuple(image_sha256),
policy=policy,
provenance=provenance,
subject=subject,
license_gate=license_gate,
)
validate_normalized_item(item.to_row())
return item
# --- ingest driver ---------------------------------------------------------
@dataclass
class IngestResult:
"""Summary of one ingest run."""
source: str
split: str
written: int
dropped: int
output_path: Path
@property
def ok(self) -> bool:
return self.written > 0 or self.dropped > 0
def write_items(
output_path: str | Path,
items: Iterable[NormalizedItem],
*,
fsync: bool = True,
) -> int:
"""Atomically write ``items`` as canonical JSONL; return the row count."""
rows = [item.to_row() for item in items]
atomic_write_jsonl(output_path, rows, fsync_dir=fsync)
return len(rows)
def append_items(output_path: str | Path, items: Iterable[NormalizedItem]) -> int:
"""Append items to an existing JSONL with per-record fsync (resume-friendly)."""
count = 0
with JsonlAppender(output_path) as appender:
for item in items:
appender.append(item.to_row())
count += 1
return count
def read_items(path: str | Path) -> Iterator[dict[str, Any]]:
"""Yield parsed normalized-item rows from ``path``."""
yield from read_jsonl(path)
# --- evaluation registry ---------------------------------------------------
def registry_row(item: NormalizedItem) -> dict[str, Any]:
"""Build one ``evaluation_items.v1.jsonl`` row from a normalized item.
Fingerprint fields (``image_phash``, ``ocr_minhash_ref``,
``question_minhash_ref``) are left ``None`` here and populated by the P3
``fingerprint`` stage; the identity + content fields are frozen now.
"""
return {
"base_id": item.base_id,
"source": item.source,
"source_revision": item.source_revision,
"config": item.source_config,
"split": item.source_split,
"native_id": item.source_native_id,
"question_sha256": item.question_sha256,
"choices_sha256": item.choices_sha256,
"image_sha256": list(item.image_sha256),
"image_phash": None,
"ocr_minhash_ref": None,
"question_minhash_ref": None,
"policy": item.policy,
}
def text_registry_row(
*,
source: str,
source_revision: str,
config: str,
split: str,
native_id: str,
question: str,
choices: Sequence[Choice] | Sequence[Mapping[str, str]],
answer_canonical: str | int | bool,
policy: Policy = "untouched_evaluation_only",
) -> dict[str, Any]:
"""Build a registry row for a text-only eval item (no image, e.g. MMLU-Pro).
Text-only ``untouched`` probes are retention checks, not intervention
candidates, so they carry no image fingerprints. They are recorded directly
in the registry rather than as visual :class:`NormalizedItem` rows.
"""
norm_choices = tuple(
c if isinstance(c, Choice) else Choice(key=str(c["key"]), text=str(c["text"]))
for c in choices
)
q_sha = compute_question_sha256(question)
c_sha = choices_sha256([{"key": c.key, "text": c.text} for c in norm_choices])
# Text-only probes carry no images: their image family is the empty set.
fam_sha = compute_image_family_sha256([])
return {
"base_id": compute_base_id(
source=source,
source_revision=source_revision,
source_native_id=native_id,
image_family_sha256=fam_sha,
question_sha256=q_sha,
choices_sha256=c_sha,
),
"source": source,
"source_revision": source_revision,
"config": config,
"split": split,
"native_id": str(native_id),
"question_sha256": q_sha,
"choices_sha256": c_sha,
"image_sha256": [],
"image_phash": None,
"ocr_minhash_ref": None,
"question_minhash_ref": None,
"policy": policy,
"answer_canonical": answer_canonical,
}
@dataclass(frozen=True)
class FrozenRegistry:
"""A write-once evaluation registry on disk."""
path: Path
sha256: str
row_count: int
def _registry_rows_keyed(rows: Iterable[Mapping[str, Any]]) -> dict[str, dict[str, Any]]:
return {str(r["base_id"]): dict(r) for r in rows}
def freeze_registry(
output_path: str | Path,
rows: Sequence[Mapping[str, Any]],
*,
force: bool = False,
resume: bool = False,
) -> FrozenRegistry:
"""Write the evaluation registry write-once; return its hash and row count.
If the registry already exists:
- ``force`` overwrites it atomically.
- ``resume`` succeeds only if the existing rows are byte-identical to the
computed rows (idempotent re-freeze); a mismatch raises
:class:`RegistryFrozenError`.
- otherwise the existing file is left untouched and a
:class:`RegistryFrozenError` is raised.
"""
path = Path(output_path)
path.parent.mkdir(parents=True, exist_ok=True)
if path.exists():
if not force and not resume:
raise RegistryFrozenError(
f"evaluation registry already frozen at {path}; "
"use --force to overwrite or --resume to verify"
)
if resume:
existing = _registry_rows_keyed(read_jsonl(path))
computed = _registry_rows_keyed(rows)
if existing != computed:
raise RegistryFrozenError(
f"evaluation registry at {path} differs from the computed "
"freeze; refusing to overwrite a frozen registry"
)
return _hash_registry(path)
atomic_write_jsonl(path, rows)
return _hash_registry(path)
def _hash_registry(path: str | Path) -> FrozenRegistry:
from ..hashing import sha256_file
rows = list(read_jsonl(path))
return FrozenRegistry(path=Path(path), sha256=sha256_file(path), row_count=len(rows))
# --- helpers shared by importers -------------------------------------------
def infer_open_answer_type(answer: str) -> AnswerType:
"""Classify an open-ended answer string as integer/number/short_text."""
text = str(answer).strip()
if text.lstrip("-+").isdigit():
return "integer"
try:
float(text)
except ValueError:
return "short_text"
return "number"
def mc_choices(texts: Sequence[str], *, keys: Sequence[str] | None = None) -> list[Choice]:
"""Build lettered choices (A, B, C, ...) from option texts in source order."""
if keys is None:
keys = [chr(ord("A") + i) for i in range(len(texts))]
if len(keys) != len(texts):
raise IngestError(f"choices/keys length mismatch: {len(keys)} keys vs {len(texts)} texts")
return [Choice(key=k, text=str(t)) for k, t in zip(keys, texts, strict=True)]
|