Datasets:
File size: 16,525 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 | """PlotQA C1 source adapter (``docs/02`` §5).
PlotQA (``NiteshMethani/PlotQA`` @ ``e0f5c34``) ships scientific plots with
crowd-sourced QA instantiated from templates. The pinned GitHub revision carries
only docs/template definitions; the **data** (plot PNGs, ``annotations.json``,
``qa_pairs_v2.json``) is hosted on Google Drive, so the git revision is not a
data pin — SHA-256 of every downloaded archive is the integrity anchor and is
recorded in the ingest manifest.
Join key: ``image_index`` (qa pair ↔ plot annotation). Only rows whose template
is in the supported allowlist and whose referenced series/x label resolves
uniquely become candidates (§5.1); executor-tolerance answer reproduction is a
P3 gate. The template→PlotDSL mapping (§5.2) is dataset-level (per-template, not
per-example).
P1 scope: deterministic ingest + ``normalize`` to a schema-conforming
:class:`NormalizedItem`, a structural ``plot_world_v1`` builder, the released
``official_answer``, and the template allowlist + mapping table. The ≥50/template
reproduction unit gate is P3 (executor).
.. note:: The exact annotation/qa field shapes are confirmed against the real
download at fetch time; the adapter is robust to the documented fields and
unit-tested with fixtures that pin the schema it relies on.
"""
from __future__ import annotations
import json
from collections.abc import Iterator, Mapping
from pathlib import Path
from typing import Any
from ..dsl.ast import Program
from ..hashing import sha256_bytes
from ..ingest.base import (
AnswerType,
ImageStore,
IngestError,
NormalizedItem,
Policy,
infer_open_answer_type,
make_item,
mc_choices,
)
from .base import (
AdapterError,
CertificateTier,
RawItem,
World,
extract_zip,
gdrive_download,
store_images,
)
SOURCE = "plotqa"
WORLD_SCHEMA = "plot_world_v1"
_VALID_SPLITS = frozenset({"train", "validation", "test"})
_MC_KEYS = ("A", "B", "C", "D", "E", "F", "G", "H")
# Supported template allowlist (§5.1). Templates not listed here are dropped
# wholesale — never patched per-row (§5.2). Extended as mappings are validated.
TEMPLATE_ALLOWLIST: frozenset[str] = frozenset(
{"T_LOOKUP", "T_DIFFERENCE", "T_MAX_YEAR", "T_MIN_YEAR", "T_SUM"}
)
# Dataset-level template → PlotDSL mapping (§5.2). Per-template, not per-example.
# The ≥50-official-reproductions-per-template gate is a P3 executor test.
TEMPLATE_DSL_MAPPING: dict[str, str] = {
"T_LOOKUP": "LOOKUP(SELECT_SERIES({series}), SELECT_X({x}))",
"T_DIFFERENCE": "SUBTRACT(LOOKUP({series}, {x1}), LOOKUP({series}, {x2}))",
"T_MAX_YEAR": "ARGMAX(POINTS(SELECT_SERIES({series})), value)",
"T_MIN_YEAR": "ARGMIN(POINTS(SELECT_SERIES({series})), value)",
"T_SUM": "SUM(POINTS(SELECT_SERIES({series})))",
}
# Google-Drive file ids per split (image archives). Annotation/qa-pair file ids
# are supplied via ``fetch``'s ``file_ids`` map; an unknown id is a hard error,
# never a silent skip. Image ids come from PlotQA_Dataset.md at the pinned rev.
_GDRIVE_IMAGE_IDS: dict[str, str] = {
"train": "1AYuaPX-Lx7T0GZvnsPgN11Twq2FZbWXL",
"validation": "1i74NRCEb-x44xqzAovuglex5d583qeiF",
"test": "1D_WPUy91vOrFl6cJUkE55n3ZuB6Qrc4u",
}
class PlotQAAdapter:
"""C1 adapter for PlotQA."""
def __init__(
self,
raw_dir: Path,
store: ImageStore,
*,
revision: str,
images_dir: Path | None = None,
) -> None:
self.raw_dir = Path(raw_dir)
self.store = store
self.revision = revision
# Plot images live under ``<raw_dir>/png/<index>.png`` by default; the
# constructor lets a test pin an explicit images directory.
self.images_dir = images_dir or (self.raw_dir / "png")
# --- fetch -----------------------------------------------------------
@classmethod
def fetch(
cls,
raw_dir: Path,
split: str,
*,
file_ids: Mapping[str, str],
expected_sha256: Mapping[str, str] | None = None,
) -> Path:
"""Download one split's archives from Google Drive into ``raw_dir``.
``file_ids`` maps a local archive name (e.g. ``"train_images.zip"``,
``"train_annotations.json"``, ``"train_qa_pairs_v2.json"``) to a Drive
file id. An id that is not supplied for a required artifact raises
:class:`AdapterError` — never a silent skip. The image archive is
extracted in place so ``_image_bytes`` can read ``png/<index>.png``.
"""
if split not in _VALID_SPLITS:
raise AdapterError(f"{SOURCE}: unsupported split {split!r}")
raw_dir = Path(raw_dir)
raw_dir.mkdir(parents=True, exist_ok=True)
required = (f"{split}_images.zip", f"{split}_annotations.json", f"{split}_qa_pairs_v2.json")
for archive in required:
file_id = file_ids.get(archive)
if not file_id:
raise AdapterError(
f"{SOURCE}: missing Drive file id for {archive!r}; "
"supply it in resources/structured_sources.plotqa.file_ids"
)
dst = raw_dir / archive
digest = (expected_sha256 or {}).get(archive)
gdrive_download(file_id, dst, expected_sha256=digest)
if archive.endswith(".zip"):
extract_zip(dst, raw_dir)
return raw_dir
@classmethod
def is_materialized(cls, raw_dir: Path, split: str) -> bool:
"""True when one split's annotation + qa files are already on disk."""
raw_dir = Path(raw_dir)
return (raw_dir / f"{split}_qa_pairs_v2.json").exists() and (
raw_dir / f"{split}_annotations.json"
).exists()
@classmethod
def materialize(
cls,
raw_dir: Path,
split: str,
*,
source_config: Mapping[str, Any],
expected_sha256: Mapping[str, str] | None = None,
) -> Path:
"""Fetch a split from Drive using file ids from the resources config.
The Drive file ids are not git-pinned, so they (and the per-archive
sha256) must be supplied in ``source_config["file_ids"]``; their absence
is a hard :class:`AdapterError`, never a silent skip.
"""
file_ids = source_config.get("file_ids") or {}
if not isinstance(file_ids, Mapping):
raise AdapterError(f"{SOURCE}: file_ids must be a mapping in resources config")
return cls.fetch(raw_dir, split, file_ids=file_ids, expected_sha256=expected_sha256)
# --- iter ------------------------------------------------------------
def _annotation_index(self) -> dict[int, dict[str, Any]]:
path = self._first_split_file("annotations.json")
records = json.loads(path.read_text(encoding="utf-8"))
out: dict[int, dict[str, Any]] = {}
for rec in records:
idx = rec.get("image_index") if isinstance(rec, dict) else None
if idx is None:
continue
out[int(idx)] = rec
return out
def _qa_pairs(self, split: str) -> list[dict[str, Any]]:
path = self._first_split_file("qa_pairs_v2.json")
data = json.loads(path.read_text(encoding="utf-8"))
return [d for d in data if isinstance(d, dict)]
def _first_split_file(self, suffix: str) -> Path:
for prefix in ("train", "validation", "test"):
candidate = self.raw_dir / f"{prefix}_{suffix}"
if candidate.exists():
return candidate
raise AdapterError(f"{SOURCE}: missing {suffix} under {self.raw_dir}")
def _image_bytes(self, image_index: int) -> bytes:
for candidate in (
self.images_dir / f"{image_index}.png",
self.raw_dir / "png" / f"{image_index}.png",
):
if candidate.exists():
return candidate.read_bytes()
raise AdapterError(f"{SOURCE}: plot image not found for image_index {image_index}")
def iter_base_items(self, split: str) -> Iterator[RawItem]:
if split not in _VALID_SPLITS:
raise AdapterError(f"{SOURCE}: unsupported split {split!r}")
annotations = self._annotation_index()
qa_pairs = self._qa_pairs(split)
# Deterministic order: (image_index, position in qa file).
ordered = sorted(
enumerate(qa_pairs),
key=lambda pair: (
int(pair[1].get("image_index", -1)) if isinstance(pair[1], dict) else -1,
pair[0],
),
)
for qa_index, qa in ordered:
if not isinstance(qa, dict):
continue
template = qa.get("template")
if not isinstance(template, str) or template not in TEMPLATE_ALLOWLIST:
continue # out-of-allowlist template → whole-template skip (§5.2)
image_index = qa.get("image_index")
if image_index is None or int(image_index) not in annotations:
continue # referenced plot missing → not a candidate (§5.1)
annotation = annotations[int(image_index)]
if not _single_series_x_match(qa, annotation):
continue # referenced series/x not unique → not a candidate (§5.1)
try:
images = {"plot": self._image_bytes(int(image_index))}
except AdapterError:
continue # image unavailable → drop, not a hard error at filter stage
payload = {
"image_index": int(image_index),
"qa_index": qa_index,
**qa,
"annotation": annotation,
}
yield RawItem(
source=SOURCE,
split=split,
source_revision=self.revision,
native_id=f"{image_index}:{qa_index}",
payload=payload,
images=images,
)
# --- normalize -------------------------------------------------------
def normalize(self, raw: RawItem) -> NormalizedItem:
qa = raw.payload
question = qa.get("question_string")
if not isinstance(question, str) or not question.strip():
raise IngestError(f"{SOURCE}/{raw.native_id}: question_string missing/empty")
answer = qa.get("answer")
if answer is None or (isinstance(answer, str) and not answer.strip()):
raise IngestError(f"{SOURCE}/{raw.native_id}: answer missing/empty")
if not raw.images:
raise IngestError(f"{SOURCE}/{raw.native_id}: plot image missing")
paths, shas = store_images(self.store, raw.images)
choices_texts = qa.get("choices")
answer_type: AnswerType
if isinstance(choices_texts, list) and choices_texts:
keys = list(_MC_KEYS[: len(choices_texts)])
choices = mc_choices([str(c) for c in choices_texts], keys=keys)
answer_type = "multiple_choice"
answer_canonical = str(answer)
else:
choices = []
answer_type = infer_open_answer_type(str(answer))
answer_canonical = str(answer)
annotation = qa.get("annotation") or {}
policy: Policy = (
"c1_train_candidate"
if raw.split in ("train", "validation")
else "c1_certified_eval_candidate"
)
extra = {
"template": qa.get("template"),
"type": qa.get("type"),
"models": annotation.get("models"),
"axes": {
"x_axis": annotation.get("x_axis"),
"y_axis": annotation.get("y_axis"),
"legend": annotation.get("legend"),
"title": annotation.get("title"),
},
# QA slot metadata consumed by the dataset-level program compiler
# (§5.2). Carrying source-native slots is not per-example human
# labeling; it is the template's structured input.
"slots": {
"series": qa.get("series") or qa.get("series_label"),
"x": qa.get("x"),
"x1": qa.get("x1"),
"x2": qa.get("x2"),
},
"raw_payload_sha256": sha256_bytes(
json.dumps(qa, sort_keys=True, separators=(",", ":")).encode("utf-8")
),
}
return make_item(
source=SOURCE,
source_revision=raw.source_revision,
source_config="default",
source_split=raw.split,
source_native_id=raw.native_id,
question=question,
choices=choices,
answer_raw=str(answer),
answer_canonical=answer_canonical,
answer_type=answer_type,
image_paths=paths,
image_sha256=shas,
policy=policy,
native_row=qa,
extra_provenance=extra,
)
# --- world / answer / tier ------------------------------------------
def build_world(self, item: NormalizedItem) -> World:
prov = item.provenance
models = prov.get("models") or []
axes = prov.get("axes") or {}
series = _series_from_models(models, axes)
return {
"world_schema": WORLD_SCHEMA,
"plot_type": _infer_plot_type(models, axes),
"series": series,
"x_axis": axes.get("x_axis"),
"y_axis": axes.get("y_axis"),
"legend": axes.get("legend"),
"title": axes.get("title"),
"style_seed": 0,
"node_visibility": {},
"provenance": {
"source": item.source,
"source_native_id": item.source_native_id,
"template": prov.get("template"),
},
}
def get_or_compile_program(self, item: NormalizedItem) -> Program:
"""Compile the typed ``plotqa_dsl_v1`` program for ``item`` (P2).
Builds the world, hands it to the dataset-level compiler, and returns a
:class:`Program` (``compiled`` or ``unsupported``). No execution, no
model calls — those are P3/P4.
"""
from ..dsl.plotdsl import compile_plotqa
return compile_plotqa(item, self.build_world(item))
def official_answer(self, item: NormalizedItem) -> str:
return str(item.answer_canonical)
def source_certificate_tier(self, item: NormalizedItem) -> CertificateTier:
return "C1_SOURCE_NATIVE"
# --- candidate-filter + world helpers -------------------------------------
def _single_series_x_match(qa: Mapping[str, Any], annotation: Mapping[str, Any]) -> bool:
"""§5.1: the referenced series/x label must resolve to exactly one in the plot.
A best-effort structural check that the QA's referenced series label (if
extractable from the template slots) matches exactly one model in the
annotation. The full slot/template-metadata consistency check is completed
with the executor (P3); here we reject only the clearly-ambiguous case.
"""
models = annotation.get("models") or []
if not isinstance(models, list) or not models:
return False
series = qa.get("series") or qa.get("series_label")
if series is None:
return True # no series slot to check (e.g. aggregate templates)
matches = [m for m in models if isinstance(m, dict) and m.get("label") == series]
return len(matches) == 1
def _series_from_models(models: Any, axes: Mapping[str, Any]) -> list[dict[str, Any]]:
if not isinstance(models, list):
return []
series: list[dict[str, Any]] = []
for i, model in enumerate(models):
if not isinstance(model, dict):
continue
points = [
{"id": f"point:{i}:{j}", "x": str(p.get("x", "")), "y": str(p.get("y", ""))}
for j, p in enumerate(model.get("points") or [])
if isinstance(p, dict)
]
series.append(
{
"id": f"series:{i}",
"label": model.get("label") or model.get("name") or f"series_{i}",
"points": points,
}
)
return series
def _infer_plot_type(models: Any, axes: Mapping[str, Any]) -> str:
legend = axes.get("legend")
if isinstance(models, list) and len(models) > 1:
return "multi_line" if legend else "bar"
return "line"
__all__ = [
"PlotQAAdapter",
"SOURCE",
"TEMPLATE_ALLOWLIST",
"TEMPLATE_DSL_MAPPING",
"WORLD_SCHEMA",
]
|