Datasets:
File size: 10,871 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 | """Pinned Hugging Face materialization for untouched retention benchmarks.
The legacy freeze path consumed hand-exported ``native rows`` whose fixture
shape did not match the real repositories (MMMU-Pro has ``image_1..image_7``
and stringified options; MathVista uses ``pid``; MMLU-Pro uses
``question_id``). This module loads the exact revisions in resources.yaml and
normalizes the real schemas directly into the common evaluation manifest.
"""
from __future__ import annotations
import ast
import io
from collections import Counter
from collections.abc import Callable, Mapping, Sequence
from pathlib import Path
from typing import Any, cast
from ..atomic_io import atomic_write_bytes, atomic_write_jsonl
from ..config import ResourcesManifest
from ..hashing import canonical_json_hash
from ..ingest.base import canonicalize_mc_answer, infer_open_answer_type, mc_choices
from .core import (
EvaluationError,
flatten_retention_items,
load_evaluation_manifest,
)
RETENTION_SOURCES: tuple[str, ...] = (
"mmmu_pro",
"mathvision",
"mathvista",
"mmlu_pro_text",
)
_SOURCE_SPECS: dict[str, tuple[str, str, int]] = {
"mmmu_pro": ("standard (10 options)", "test", 1_730),
"mathvision": ("default", "testmini", 304),
"mathvista": ("default", "testmini", 1_000),
"mmlu_pro_text": ("default", "test", 12_032),
}
DatasetLoader = Callable[..., Sequence[Mapping[str, Any]]]
def _load_dataset(
repo_id: str,
config: str,
*,
split: str,
revision: str,
cache_dir: str | None,
) -> Sequence[Mapping[str, Any]]:
try:
from datasets import load_dataset
except ImportError as exc:
raise EvaluationError(f"datasets is required for retention download: {exc}") from exc
return cast(
Sequence[Mapping[str, Any]],
load_dataset(
repo_id,
config,
split=split,
revision=revision,
cache_dir=cache_dir,
),
)
def _options(value: Any, *, source: str, native_id: str) -> list[str]:
if value in (None, ""):
return []
parsed = value
if isinstance(value, str):
try:
parsed = ast.literal_eval(value)
except (SyntaxError, ValueError) as exc:
raise EvaluationError(
f"{source}/{native_id}: options string is not a Python list literal"
) from exc
if not isinstance(parsed, Sequence) or isinstance(parsed, str | bytes | bytearray):
raise EvaluationError(f"{source}/{native_id}: options must be a sequence")
result = [str(option) for option in parsed]
if any(not option for option in result):
raise EvaluationError(f"{source}/{native_id}: options contain an empty value")
return result
def _native_id(source: str, row: Mapping[str, Any]) -> str:
field = {
"mmmu_pro": "id",
"mathvision": "id",
"mathvista": "pid",
"mmlu_pro_text": "question_id",
}[source]
value = row.get(field)
if value is None or not str(value):
raise EvaluationError(f"{source}: row has no {field}")
return str(value)
def _image_values(source: str, row: Mapping[str, Any]) -> list[Any]:
if source == "mmmu_pro":
return [row.get(f"image_{index}") for index in range(1, 8) if row.get(f"image_{index}")]
if source in {"mathvision", "mathvista"}:
value = row.get("decoded_image")
return [value] if value is not None else []
return []
def _png_bytes(value: Any, *, label: str) -> bytes:
try:
from PIL import Image, ImageOps
except ImportError as exc:
raise EvaluationError(f"Pillow is required for retention images: {exc}") from exc
image: Any
try:
if isinstance(value, Image.Image):
image = value.copy()
elif isinstance(value, Mapping) and isinstance(value.get("bytes"), bytes):
image = Image.open(io.BytesIO(value["bytes"])).copy()
elif isinstance(value, Mapping) and isinstance(value.get("path"), str):
image = Image.open(str(value["path"])).copy()
elif isinstance(value, str | Path):
image = Image.open(str(value)).copy()
else:
raise EvaluationError(f"{label}: unsupported decoded image value")
image = ImageOps.exif_transpose(image).convert("RGB")
buffer = io.BytesIO()
image.save(buffer, format="PNG", optimize=False)
return buffer.getvalue()
except (OSError, ValueError) as exc:
raise EvaluationError(f"{label}: cannot decode image: {exc}") from exc
def _store_images(
source: str,
row_index: int,
native_id: str,
values: Sequence[Any],
asset_root: Path,
) -> list[str]:
if source != "mmlu_pro_text" and not values:
raise EvaluationError(f"{source}/{native_id}: visual retention row has no image")
paths: list[str] = []
for image_index, value in enumerate(values):
relative = Path(source) / f"{row_index:06d}" / f"image-{image_index}.png"
atomic_write_bytes(
asset_root / relative,
_png_bytes(value, label=f"{source}/{native_id} image {image_index}"),
fsync_dir=False,
)
paths.append(relative.as_posix())
return paths
def _item(
source: str,
row: Mapping[str, Any],
*,
row_index: int,
revision: str,
config: str,
split: str,
asset_root: Path,
) -> dict[str, Any]:
native_id = _native_id(source, row)
question = row.get("question")
answer = row.get("answer")
if not isinstance(question, str) or not question.strip():
raise EvaluationError(f"{source}/{native_id}: question is empty")
if answer is None or not str(answer).strip():
raise EvaluationError(f"{source}/{native_id}: answer is empty")
raw_options = _options(
row.get("options") if source != "mathvista" else row.get("choices"),
source=source,
native_id=native_id,
)
choices = [choice.to_dict() for choice in mc_choices(raw_options)]
if choices:
answer_type = "multiple_choice"
canonical_answer = canonicalize_mc_answer(str(answer), mc_choices(raw_options))
else:
answer_type = infer_open_answer_type(str(answer))
canonical_answer = str(answer)
image_paths = _store_images(
source,
row_index,
native_id,
_image_values(source, row),
asset_root,
)
return {
"base_id": canonical_json_hash(
{
"source": source,
"revision": revision,
"config": config,
"split": split,
"native_id": native_id,
}
),
"source": source,
"source_revision": revision,
"source_config": config,
"source_split": split,
"question": question,
"choices": choices,
"image_paths": image_paths,
"answer_type": answer_type,
"answer_canonical": canonical_answer,
"subject": str(row.get("subject") or row.get("category") or "unknown"),
}
def _existing_summary(
output_path: Path,
asset_root: Path,
*,
expected_sources: Sequence[str],
) -> dict[str, Any]:
rows = load_evaluation_manifest(output_path)
for row in rows:
images = row.get("images")
if not isinstance(images, list):
raise EvaluationError("existing retention row images are malformed")
for image in images:
if not isinstance(image, Mapping) or not isinstance(image.get("path"), str):
raise EvaluationError("existing retention image record is malformed")
if not (asset_root / str(image["path"])).is_file():
raise EvaluationError(f"existing retention image is missing: {image['path']}")
source_counts = Counter(str(row["source"]) for row in rows)
if set(source_counts) != set(expected_sources):
raise EvaluationError(
"existing retention manifest sources differ from the requested source set"
)
return {
"schema_version": 1,
"kind": "pinned_hf_retention_evaluation_manifest",
"status": "reused",
"path": str(output_path.resolve()),
"item_count": len(rows),
"source_counts": dict(sorted(source_counts.items())),
}
def materialize_hf_retention(
resources: ResourcesManifest,
output_path: Path,
asset_root: Path,
*,
cache_dir: Path | None = None,
sources: Sequence[str] = RETENTION_SOURCES,
dataset_loader: DatasetLoader = _load_dataset,
expected_counts: Mapping[str, int] | None = None,
) -> dict[str, Any]:
"""Download pinned untouched sources and freeze the common eval manifest."""
selected = tuple(dict.fromkeys(sources))
if not selected or any(source not in RETENTION_SOURCES for source in selected):
raise EvaluationError(f"retention sources must come from {RETENTION_SOURCES}")
if output_path.exists():
return _existing_summary(output_path, asset_root, expected_sources=selected)
asset_root.mkdir(parents=True, exist_ok=True)
items: list[dict[str, Any]] = []
counts: dict[str, int] = {}
for source in selected:
config, split, registered_count = _SOURCE_SPECS[source]
resource = resources.dataset(source)
if resource.role not in {"untouched_eval_only", "untouched_text_reasoning_retention_eval"}:
raise EvaluationError(f"{source} is not registered as untouched evaluation")
rows = dataset_loader(
resource.repo_id,
config,
split=split,
revision=resource.revision,
cache_dir=str(cache_dir) if cache_dir is not None else None,
)
count = len(rows)
wanted = (
expected_counts[source]
if expected_counts is not None and source in expected_counts
else registered_count
)
if count != wanted:
raise EvaluationError(f"{source}: pinned split has {count} rows, expected {wanted}")
counts[source] = count
items.extend(
_item(
source,
row,
row_index=index,
revision=resource.revision,
config=config,
split=split,
asset_root=asset_root,
)
for index, row in enumerate(rows)
)
evaluation_rows = flatten_retention_items(items)
atomic_write_jsonl(output_path, evaluation_rows)
return {
"schema_version": 1,
"kind": "pinned_hf_retention_evaluation_manifest",
"status": "completed",
"path": str(output_path.resolve()),
"asset_root": str(asset_root.resolve()),
"item_count": len(evaluation_rows),
"source_counts": dict(sorted(counts.items())),
}
|