Datasets:
File size: 15,103 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 | """Geometry3K C1 source adapter (``docs/02`` §6).
Geometry3K (``lupantech/InterGPS`` @ ``99e6b52``) ships 3,002 high-school
geometry problems across ``train`` / ``validation`` / ``test``. Each problem has
a released ``data.json`` (problem text, four choices, answer letter, diagram
image) and an **annotated** logic form; the global ``logic_forms.zip`` holds the
authoritative annotated text/diagram logic forms keyed by integer ``id``. C1 uses
the annotated forms only — predicted diagram-parser output is excluded
(``docs/02`` §6.1).
P1 scope: deterministic ingest + ``normalize`` to a schema-conforming
:class:`NormalizedItem`, a structural ``geometry_world_v1`` builder with channel
provenance, and the released ``official_answer``. The typed DSL AST (P2) and the
6-check solver acceptance (P3) are deferred.
"""
from __future__ import annotations
import json
import re
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 (
ImageStore,
IngestError,
NormalizedItem,
Policy,
make_item,
mc_choices,
)
from .base import (
AdapterError,
CertificateTier,
Channel,
RawItem,
World,
extract_zip,
github_raw_url,
http_download,
store_images,
)
SOURCE = "geometry3k"
WORLD_SCHEMA = "geometry_world_v1"
# The release names its split archive/directory ``val``; the resources manifest
# records the conventional ``validation``. Accept either.
_SPLIT_ALIAS: dict[str, str] = {"validation": "val"}
_VALID_SPLITS = frozenset({"train", "validation", "test"})
_MC_KEYS = ("A", "B", "C", "D")
# Archives fetched from the pinned revision. ``symbols.zip`` (predicted diagram
# symbols) is intentionally excluded — C1 uses annotated forms only (§6.1).
_ARCHIVES = ("train.zip", "val.zip", "test.zip", "logic_forms.zip")
# Annotated logic-form files inside ``logic_forms.zip``, keyed by integer id.
_TEXT_LOGIC_FORMS = "text_logic_forms_annot.json"
_TEXT_LOGIC_FORMS_DISSOLVED = "text_logic_forms_annot_dissolved.json"
_DIAGRAM_LOGIC_FORMS = "diagram_logic_forms_annot.json"
class Geometry3KAdapter:
"""C1 adapter for Geometry3K."""
def __init__(self, raw_dir: Path, store: ImageStore, *, revision: str) -> None:
self.raw_dir = Path(raw_dir)
self.store = store
self.revision = revision
self._logic_forms: dict[str, dict[int, Any]] | None = None
# --- fetch -----------------------------------------------------------
@classmethod
def fetch(
cls,
raw_dir: Path,
*,
repo_id: str,
revision: str,
data_path: str,
expected_sha256: Mapping[str, str] | None = None,
) -> Path:
"""Download + extract the Geometry3K archives into ``raw_dir``.
Idempotent: existing verified archives are kept. ``expected_sha256`` maps
archive name to a known-good digest when available; absent entries
download without verification (the git revision is the pin) but the
computed digest is still recorded by the caller's manifest.
"""
raw_dir = Path(raw_dir)
raw_dir.mkdir(parents=True, exist_ok=True)
for archive in _ARCHIVES:
url = github_raw_url(repo_id, revision, f"{data_path}/{archive}")
dst = raw_dir / archive
digest = (expected_sha256 or {}).get(archive)
http_download(url, dst, expected_sha256=digest)
extract_zip(dst, raw_dir)
return raw_dir
@classmethod
def is_materialized(cls, raw_dir: Path, split: str) -> bool:
"""True when the split dir and annotated logic forms are on disk."""
if split not in _VALID_SPLITS:
return False
raw_dir = Path(raw_dir)
dir_name = _SPLIT_ALIAS.get(split, split)
split_present = (raw_dir / dir_name).is_dir()
logic_forms_present = (raw_dir / "logic_forms").is_dir() or (
raw_dir / _TEXT_LOGIC_FORMS
).exists()
return split_present and logic_forms_present
@classmethod
def materialize(
cls,
raw_dir: Path,
split: str,
*,
source_config: Mapping[str, Any],
expected_sha256: Mapping[str, str] | None = None,
) -> Path:
"""Fetch all Geometry3K archives git-pinned via the resources config.
``source_config`` carries ``repo_id``, ``revision``, and ``data_path``
from ``resources.structured_sources.geometry3k``. Fetch grabs every split
archive (idempotent), so ``split`` only gates ``is_materialized``.
"""
repo_id = source_config.get("repo_id")
revision = source_config.get("revision")
data_path = source_config.get("data_path")
if not (
isinstance(repo_id, str) and isinstance(revision, str) and isinstance(data_path, str)
):
raise AdapterError(
f"{SOURCE}: materialize requires repo_id, revision, data_path in resources config"
)
return cls.fetch(
raw_dir,
repo_id=repo_id,
revision=revision,
data_path=data_path,
expected_sha256=expected_sha256,
)
# --- iter ------------------------------------------------------------
def _load_logic_forms(self) -> dict[str, dict[int, Any]]:
if self._logic_forms is not None:
return self._logic_forms
root = self.raw_dir / "logic_forms"
if not root.is_dir():
# Some extractions place the three json files directly under raw_dir.
root = self.raw_dir
loaded: dict[str, dict[int, Any]] = {}
for fname in (_TEXT_LOGIC_FORMS, _TEXT_LOGIC_FORMS_DISSOLVED, _DIAGRAM_LOGIC_FORMS):
path = root / fname
if not path.exists():
raise AdapterError(f"missing annotated logic-form file: {path}")
data = json.loads(path.read_text(encoding="utf-8"))
# The release keys these by integer id (stringified in JSON).
keyed = {int(k): v for k, v in data.items()}
loaded[fname] = keyed
self._logic_forms = loaded
return loaded
def iter_base_items(self, split: str) -> Iterator[RawItem]:
if split not in _VALID_SPLITS:
raise AdapterError(f"{SOURCE}: unsupported split {split!r}")
dir_name = _SPLIT_ALIAS.get(split, split)
split_dir = self.raw_dir / dir_name
if not split_dir.is_dir():
raise AdapterError(f"{SOURCE}: split directory not found: {split_dir}")
logic = self._load_logic_forms()
text_forms = logic[_TEXT_LOGIC_FORMS_DISSOLVED] or logic[_TEXT_LOGIC_FORMS]
diagram_forms = logic[_DIAGRAM_LOGIC_FORMS]
ids = sorted(int(p.name) for p in split_dir.iterdir() if p.is_dir() and p.name.isdigit())
for pid in ids:
problem_dir = split_dir / str(pid)
data_path = problem_dir / "data.json"
if not data_path.exists():
raise AdapterError(f"{SOURCE}/{pid}: missing data.json")
data = json.loads(data_path.read_text(encoding="utf-8"))
img_path = problem_dir / "img_diagram.png"
if not img_path.exists():
raise AdapterError(f"{SOURCE}/{pid}: missing img_diagram.png")
images = {"diagram": img_path.read_bytes()}
text_lf = text_forms.get(pid)
diagram_lf = diagram_forms.get(pid)
if text_lf is None or diagram_lf is None:
# A problem without annotated forms cannot be a C1 candidate.
raise AdapterError(f"{SOURCE}/{pid}: missing annotated logic forms")
payload = {
"id": pid,
**data,
"text_logic_form": text_lf,
"diagram_logic_form": diagram_lf,
}
yield RawItem(
source=SOURCE,
split=split,
source_revision=self.revision,
native_id=str(pid),
payload=payload,
images=images,
)
# --- normalize -------------------------------------------------------
def normalize(self, raw: RawItem) -> NormalizedItem:
data = raw.payload
problem_text = data.get("problem_text")
if not isinstance(problem_text, str) or not problem_text.strip():
raise IngestError(f"{SOURCE}/{raw.native_id}: problem_text missing/empty")
choice_texts = data.get("choices")
if not isinstance(choice_texts, list) or len(choice_texts) != len(_MC_KEYS):
raise IngestError(f"{SOURCE}/{raw.native_id}: choices must be {len(_MC_KEYS)} strings")
choices = mc_choices([str(c) for c in choice_texts], keys=list(_MC_KEYS))
answer = data.get("answer")
if not isinstance(answer, str) or answer not in _MC_KEYS:
raise IngestError(
f"{SOURCE}/{raw.native_id}: answer must be a letter A-D, got {answer!r}"
)
if not raw.images:
raise IngestError(f"{SOURCE}/{raw.native_id}: diagram image missing")
paths, shas = store_images(self.store, raw.images)
graph_types = data.get("problem_type_graph") or []
subject = (
str(graph_types[0]) if isinstance(graph_types, list) and graph_types else "geometry"
)
policy: Policy = (
"c1_train_candidate"
if raw.split in ("train", "validation")
else "c1_certified_eval_candidate"
)
goal = data.get("problem_type_goal")
channel_map = _channel_map(data.get("text_logic_form"), data.get("diagram_logic_form"))
extra = {
"annotated_text_logic_form": data.get("text_logic_form"),
"annotated_diagram_logic_form": data.get("diagram_logic_form"),
"channel_map": channel_map,
"goal": goal,
"problem_type_goal": goal,
"raw_payload_sha256": sha256_bytes(
json.dumps(data, 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=problem_text,
choices=choices,
answer_raw=answer,
answer_canonical=answer,
answer_type="multiple_choice",
image_paths=paths,
image_sha256=shas,
policy=policy,
native_row=data,
subject=subject,
extra_provenance=extra,
)
# --- world / answer / tier ------------------------------------------
def build_world(self, item: NormalizedItem) -> World:
prov = item.provenance
text_lf = _as_logic_lines(prov.get("annotated_text_logic_form"))
diagram_lf = _as_logic_lines(prov.get("annotated_diagram_logic_form"))
channel_map = prov.get("channel_map") or {}
entities: dict[str, dict[str, Any]] = {}
constraints: list[dict[str, Any]] = []
for line in text_lf:
pred = _parse_predicate(line)
if pred is not None:
constraints.append({**pred, "channel": channel_map.get(line, "text")})
_collect_entities(pred, entities)
for line in diagram_lf:
pred = _parse_predicate(line)
if pred is not None:
constraints.append({**pred, "channel": channel_map.get(line, "visual")})
_collect_entities(pred, entities)
return {
"world_schema": WORLD_SCHEMA,
"entities": [{"id": eid, **body} for eid, body in sorted(entities.items())],
"constraints": constraints,
"goal": prov.get("goal"),
"provenance": {
"source": item.source,
"source_native_id": item.source_native_id,
"channels": sorted(set(channel_map.values()) or {"text", "visual"}),
},
}
def get_or_compile_program(self, item: NormalizedItem) -> Program:
"""Compile the typed ``geometry_dsl_v1`` program for ``item`` (P2).
Builds the world, hands it to the compiler, and returns a
:class:`Program` (``compiled`` or ``unsupported``) with constraint
channel provenance on the envelope. No execution, no model calls — P3/P4.
"""
from ..dsl.geometrydsl import compile_geometry3k
return compile_geometry3k(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"
# --- logic-form parsing helpers -------------------------------------------
_PRED_RE = re.compile(r"^([A-Za-z_]+)\((.*)\)$")
def _as_logic_lines(value: Any) -> list[str]:
"""Normalize an annotated logic-form value to a list of predicate strings."""
if value is None:
return []
if isinstance(value, str):
return [ln.strip() for ln in value.splitlines() if ln.strip()]
if isinstance(value, list):
return [str(ln).strip() for ln in value if str(ln).strip()]
return []
def _parse_predicate(line: str) -> dict[str, Any] | None:
match = _PRED_RE.match(line.strip())
if match is None:
return None
predicate, raw_args = match.groups()
args = [a.strip() for a in raw_args.split(",")] if raw_args else []
return {"predicate": predicate, "args": args}
def _collect_entities(pred: Mapping[str, Any], entities: dict[str, dict[str, Any]]) -> None:
"""Record referenced entity ids (``point:A``, ``segment:A:B``, ...) from args."""
for arg in pred["args"]:
if isinstance(arg, str) and ":" in arg:
eid, _, body = arg.partition(":")
full = arg
etype = eid.capitalize() if eid else "Entity"
if full not in entities:
entities[full] = {"type": etype, "args": []}
if body and body not in entities[full]["args"]:
entities[full]["args"].append(body)
def _channel_map(text_lf: Any, diagram_lf: Any) -> dict[str, Channel]:
"""Tag each predicate string with its provenance channel (§6.2).
A fact present in both text and diagram is ``redundant``; text-only is
``text``; diagram-only is ``visual``. Interventions alter only ``visual``.
"""
text_set = set(_as_logic_lines(text_lf))
diagram_set = set(_as_logic_lines(diagram_lf))
channel_map: dict[str, Channel] = {}
for line in text_set & diagram_set:
channel_map[line] = "redundant"
for line in text_set - diagram_set:
channel_map[line] = "text"
for line in diagram_set - text_set:
channel_map[line] = "visual"
return channel_map
__all__ = ["Geometry3KAdapter", "SOURCE", "WORLD_SCHEMA"]
|