Datasets:
File size: 17,483 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 | """Per-source importers and the ingest / freeze-eval drivers.
The importer registry maps a resource logical name to its normalization function
(``docs/02_DATA_PIPELINE.md`` §4). Two drivers consume it:
* :func:`ingest_source` — normalize one *training* source to
``normalized/<source>/<split>/items.jsonl``. It hard-refuses MathVista (a
training-prohibited source) and skips BBox DocVQA when its dual approval
gate is not cleared, and it requires the evaluation registry to exist first.
* :func:`freeze_eval` — build and write-once freeze the evaluation registry
(``evaluation_items.v1.jsonl``) from the evaluation sources, before any
training ingest runs.
"""
from __future__ import annotations
from collections.abc import Callable, Mapping, Sequence
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from ..atomic_io import atomic_write_jsonl, read_jsonl
from ..config import ExperimentConfig, ResourcesManifest
from ..hashing import sha256_file
from ..manifests import write_dataset_manifest
from ..vcs import current_code_commit
from . import base, bbox_docvqa, chartqa, eval_sources, mmk12, mmmu, mmr1
from .base import (
DirectoryImageResolver,
FrozenRegistry,
ImageResolver,
ImageStore,
IngestError,
IngestResult,
NormalizedItem,
Policy,
RegistryFrozenError,
freeze_registry,
registry_row,
)
# (row, images, resolve, *, revision, split, config) -> NormalizedItem | None
NormalizeFunc = Callable[..., "NormalizedItem | None"]
@dataclass(frozen=True)
class ImporterSpec:
"""One registered source importer."""
name: str
normalize: NormalizeFunc
policy: Policy
TRAIN_IMPORTERS: dict[str, ImporterSpec] = {
"mmk12": ImporterSpec("mmk12", mmk12.normalize, "c2_train_candidate"),
"mmr1_rl": ImporterSpec("mmr1_rl", mmr1.normalize, "c2_train_candidate"),
"mmmu": ImporterSpec("mmmu", mmmu.normalize, "c2_train_candidate"),
"chartqa": ImporterSpec("chartqa", chartqa.normalize, "c2_train_candidate"),
"bbox_docvqa_train": ImporterSpec(
"bbox_docvqa_train", bbox_docvqa.normalize, "c2_train_candidate"
),
}
EVAL_VISUAL_IMPORTERS: dict[str, NormalizeFunc] = {
"mmmu_pro": eval_sources.normalize_mmmu_pro,
"mathvision": eval_sources.normalize_mathvision,
"mathvista": eval_sources.normalize_mathvista,
}
TEXT_EVAL_SOURCES = {"mmlu_pro_text"}
# MathVista's card prohibits training; it may only ever be frozen as eval.
FORBIDDEN_TRAIN_SOURCES = frozenset({"mathvista"})
TRAIN_SOURCES = frozenset(TRAIN_IMPORTERS)
EVAL_SOURCES = frozenset(EVAL_VISUAL_IMPORTERS) | TEXT_EVAL_SOURCES
def importer_for(name: str) -> ImporterSpec:
"""Return the train importer registered for ``name``."""
spec = TRAIN_IMPORTERS.get(name)
if spec is None:
raise IngestError(f"no train importer registered for source {name!r}")
return spec
def read_native_rows(path: str | Path) -> list[dict[str, Any]]:
"""Read native rows from a JSONL file (skipping blank lines)."""
return [dict(row) for row in read_jsonl(path)]
def ingest_source(
name: str,
split: str,
rows: Sequence[Mapping[str, Any]],
image_resolver: ImageResolver,
output_dir: str | Path,
*,
revision: str,
config: str = "default",
approval: Mapping[str, Mapping[str, Any]] | None = None,
eval_registry_path: str | Path | None = None,
require_eval_registry: bool = True,
config_sha256: str = "",
created_at: str = "",
) -> IngestResult:
"""Normalize one training source to ``<output_dir>/items.jsonl``.
Hard failures (nonzero semantics): ingesting a forbidden/eval-only source
as train (MathVista), a split the importer rejects, or an invariant
violation. BBox DocVQA with an uncleared gate is a *skip* (zero rows), not
an error — the core pipeline runs without it.
"""
if name in FORBIDDEN_TRAIN_SOURCES or name in EVAL_SOURCES:
raise IngestError(
f"source {name!r} is evaluation-only and must never be ingested as a training source"
)
spec = importer_for(name)
if name == "bbox_docvqa_train":
reason = bbox_docvqa.block_reason(approval or {})
if reason is not None:
# Disabled source: write an empty items file + manifest noting the skip.
out = Path(output_dir)
out.mkdir(parents=True, exist_ok=True)
items_path = out / "items.jsonl"
base.write_items(items_path, [])
_write_ingest_manifest(
out / "items.manifest.json",
items_path=items_path,
source=name,
split=split,
row_count=0,
dropped=0,
input_manifest_sha256=_registry_sha(eval_registry_path),
config_sha256=config_sha256,
created_at=created_at,
extra={"skipped": True, "skip_reason": reason},
)
return IngestResult(name, split, 0, 0, items_path)
if require_eval_registry and (
eval_registry_path is None or not Path(eval_registry_path).exists()
):
raise IngestError(
"evaluation registry must be frozen before any training ingest "
"(run `explicit-data freeze-eval` first, or pass --no-require-eval-registry)"
)
out = Path(output_dir)
out.mkdir(parents=True, exist_ok=True)
images = ImageStore(out)
items: list[NormalizedItem] = []
dropped = 0
for row in rows:
result = spec.normalize(
row, images, image_resolver, revision=revision, split=split, config=config
)
if result is None:
dropped += 1
continue
items.append(result)
items_path = out / "items.jsonl"
base.write_items(items_path, items)
_write_ingest_manifest(
out / "items.manifest.json",
items_path=items_path,
source=name,
split=split,
row_count=len(items),
dropped=dropped,
input_manifest_sha256=_registry_sha(eval_registry_path),
config_sha256=config_sha256,
created_at=created_at,
)
return IngestResult(name, split, len(items), dropped, items_path)
def _registry_sha(path: str | Path | None) -> str | None:
if path is None or not Path(path).exists():
return None
return sha256_file(path)
def ingest_structured_source(
name: str,
split: str,
raw_dir: str | Path,
output_dir: str | Path,
*,
revision: str,
source_config: Mapping[str, Any] | None = None,
expected_sha256: Mapping[str, str] | None = None,
eval_registry_path: str | Path | None = None,
require_eval_registry: bool = True,
config_sha256: str = "",
created_at: str = "",
) -> IngestResult:
"""Normalize one C1 structured source (PlotQA/Geometry3K) to ``items.jsonl``.
The structured path reads from a materialized ``raw_dir``. If the split's
artifacts are absent, ``adapter.materialize`` fetches them (idempotent,
sha256-verified); a blocked or missing artifact raises
:class:`~explicit_learning.sources.base.AdapterError` — never a silent skip
and never a substitute source. Train/validation splits carry
``c1_train_candidate`` and require the evaluation registry to be frozen
first (mirroring :func:`ingest_source`); the test split
(``c1_certified_eval_candidate``) is eval and skips that gate — certified-eval
is generated later by ``build-certified-eval`` (P3).
"""
from ..sources import ADAPTERS
from ..sources.base import AdapterError
if name not in ADAPTERS:
raise IngestError(f"no structured-source adapter registered for {name!r}")
adapter_cls = ADAPTERS[name]
raw_dir = Path(raw_dir)
out = Path(output_dir)
out.mkdir(parents=True, exist_ok=True)
is_train = split in ("train", "validation")
if (
is_train
and require_eval_registry
and (eval_registry_path is None or not Path(eval_registry_path).exists())
):
raise IngestError(
"evaluation registry must be frozen before any training ingest "
"(run `explicit-data freeze-eval` first, or pass --no-require-eval-registry)"
)
if not adapter_cls.is_materialized(raw_dir, split):
try:
adapter_cls.materialize(
raw_dir,
split,
source_config=source_config or {},
expected_sha256=expected_sha256,
)
except AdapterError as exc:
raise IngestError(
f"structured source {name!r} could not be materialized: {exc}"
) from exc
store = ImageStore(out)
adapter = adapter_cls(raw_dir, store, revision=revision)
items: list[NormalizedItem] = []
for raw in adapter.iter_base_items(split):
items.append(adapter.normalize(raw))
items_path = out / "items.jsonl"
base.write_items(items_path, items)
_write_ingest_manifest(
out / "items.manifest.json",
items_path=items_path,
source=name,
split=split,
row_count=len(items),
dropped=0,
input_manifest_sha256=_registry_sha(eval_registry_path) if is_train else None,
config_sha256=config_sha256,
created_at=created_at,
extra={"certificate_tier": "C1_SOURCE_NATIVE"},
)
return IngestResult(name, split, len(items), 0, items_path)
def _write_ingest_manifest(
output_path: Path,
*,
items_path: Path,
source: str,
split: str,
row_count: int,
dropped: int,
input_manifest_sha256: str | None,
config_sha256: str,
created_at: str,
extra: Mapping[str, Any] | None = None,
) -> None:
record_extra: dict[str, Any] = {
"source": source,
"split": split,
"dropped": dropped,
"input_manifest_sha256": input_manifest_sha256,
"code_commit": current_code_commit(),
"config_sha256": config_sha256 or None,
"created_at": created_at or None,
}
if extra:
record_extra.update(extra)
write_dataset_manifest(
output_path=output_path,
dataset_name=f"{source}.{split}",
items_path=items_path,
extra=record_extra,
)
# --- evaluation registry freeze -------------------------------------------
@dataclass(frozen=True)
class EvalSourceSpec:
"""One evaluation source to freeze, with its config and split.
ADR-0002 freezes only the *untouched* retention registry (MMMU-Pro,
MathVision, MathVista, MMLU-Pro) before any training ingest. The certified
intervention eval (PlotQA/Geometry3K test) is generated later by
``build-certified-eval``, not by this freeze. ``from_gold`` is retained for
callers that pin a split verbatim; the default freeze path carries an
implicit ``test`` placeholder that the caller resolves to the source's real
freeze split via :func:`default_split_for`.
"""
name: str
config: str
split: str
from_gold: bool = False
def eval_freeze_plan(experiment: ExperimentConfig) -> list[EvalSourceSpec]:
"""Build the ordered, de-duplicated list of untouched evaluation sources to freeze.
ADR-0002 freezes only the *untouched* retention registry (MMMU-Pro,
MathVision, MathVista, MMLU-Pro) before any training ingest. The certified
intervention eval (PlotQA/Geometry3K test) is generated later by
``build-certified-eval``, not by this freeze. Each spec's real freeze split
is resolved by the caller via :func:`default_split_for`.
"""
specs: dict[str, EvalSourceSpec] = {}
for name in experiment.data.evaluation.get("untouched", []):
# Untouched text/visual probe: default split resolved by the caller.
specs[name] = EvalSourceSpec(name=name, config="default", split="test")
return list(specs.values())
def default_split_for(name: str, resources: ResourcesManifest) -> str:
"""Pick the freeze split for an untouched source from its resource metadata."""
if name not in resources.datasets:
return "test"
splits = resources.datasets[name].splits or {}
for candidate in ("testmini", "test", "validation"):
if candidate in splits:
return candidate
# No conventional eval split key present (e.g. MMMU-Pro records a per-config
# count under "test_per_config" rather than a real HF split name). The
# canonical untouched freeze split is "test"; never return a count-style
# pseudo-key, which the importers would reject.
return "test"
def freeze_eval(
plan: Sequence[EvalSourceSpec],
resources: ResourcesManifest,
*,
rows_dir: str | Path,
image_root: str | Path | None,
output_path: str | Path,
force: bool = False,
resume: bool = False,
) -> FrozenRegistry:
"""Build and write-once freeze the evaluation registry.
For each eval source, native rows are read from
``<rows_dir>/<source>.<split>.jsonl``. Visual sources go through their
importer; text MMLU-Pro is recorded as a text registry row. The combined,
base_id-sorted rows are written once.
A companion ``<registry>.items.jsonl`` of the *full* normalized eval items
(text + ``image_paths``) is written beside the write-once registry so the P3
``fingerprint`` stage can compute text and pixel fingerprints for eval — the
frozen registry itself carries only hashes, not the text needed for MinHash.
Eval images are content-addressed under ``<registry_dir>/eval_images`` so a
single image root serves every eval source.
"""
rows_dir = Path(rows_dir)
registry_path = Path(output_path)
eval_images_dir = registry_path.parent / "eval_images"
resolver = DirectoryImageResolver(image_root) if image_root else None
store = ImageStore(eval_images_dir)
registry_rows: list[dict[str, Any]] = []
eval_item_rows: list[dict[str, Any]] = []
for spec in plan:
revision = resources.datasets[spec.name].revision
rows_path = rows_dir / f"{spec.name}.{spec.split}.jsonl"
if not rows_path.exists():
raise IngestError(f"missing native rows for eval source {spec.name!r}: {rows_path}")
rows = read_native_rows(rows_path)
if spec.name in TEXT_EVAL_SOURCES:
for row in rows:
reg = eval_sources.registry_row_mmlu_pro_text(
row, revision=revision, split=spec.split, config=spec.config
)
registry_rows.append(reg)
choices = base.mc_choices([str(o) for o in row["options"]])
eval_item_rows.append(
{
"schema_version": base.SCHEMA_VERSION,
"base_id": reg["base_id"],
"source": reg["source"],
"source_revision": reg["source_revision"],
"source_config": reg["config"],
"source_split": reg["split"],
"source_native_id": reg["native_id"],
"question": str(row["question"]),
"choices": [c.to_dict() for c in choices],
"choices_sha256": reg["choices_sha256"],
"question_sha256": reg["question_sha256"],
"image_paths": [],
"image_sha256": [],
"answer_raw": str(row["answer"]),
"answer_canonical": reg["answer_canonical"],
"answer_type": "multiple_choice",
"policy": reg["policy"],
"provenance": {},
"subject": str(row.get("subject") or "unknown"),
"license_gate": None,
}
)
continue
normalize = EVAL_VISUAL_IMPORTERS[spec.name]
if resolver is None:
raise IngestError(f"visual eval source {spec.name!r} requires --image-root")
for row in rows:
item = normalize(
row, store, resolver, revision=revision, split=spec.split, config=spec.config
)
if item is None:
raise IngestError(
f"eval importer for {spec.name!r} dropped a row; eval sources "
"must never be filtered at freeze time"
)
registry_rows.append(registry_row(item))
eval_item_rows.append(item.to_row())
registry_rows.sort(key=lambda r: str(r["base_id"]))
frozen = freeze_registry(output_path, registry_rows, force=force, resume=resume)
# Companion items file (regenerable, not write-once) for the fingerprint stage.
companion_path = registry_path.with_name(registry_path.stem + ".items.jsonl")
eval_item_rows.sort(key=lambda r: str(r["base_id"]))
atomic_write_jsonl(companion_path, eval_item_rows)
return frozen
__all__ = [
"EvalSourceSpec",
"ImporterSpec",
"ImageStore",
"IngestError",
"IngestResult",
"NormalizedItem",
"RegistryFrozenError",
"EVAL_SOURCES",
"EVAL_VISUAL_IMPORTERS",
"FORBIDDEN_TRAIN_SOURCES",
"TEXT_EVAL_SOURCES",
"TRAIN_IMPORTERS",
"TRAIN_SOURCES",
"base",
"default_split_for",
"eval_freeze_plan",
"freeze_eval",
"importer_for",
"ingest_source",
"ingest_structured_source",
"read_native_rows",
]
|