Datasets:
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
100K - 1M
ArXiv:
Tags:
multi-hop-question-answering
hotpotqa
evidence-selection
question-decomposition
chain-of-thought
supervised-fine-tuning
License:
File size: 13,950 Bytes
93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e 7f3a1d4 93a180e | 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 | #!/usr/bin/env python3
"""Build every verified Bactrainus train configuration as Parquet.
By default the script downloads the complete, revision-pinned official
``hotpotqa/hotpot_qa`` distractor training split. A complete official JSON or
JSONL export may be supplied instead. The builder refuses partial inputs,
writes into private staging, validates the result, and only then installs it.
"""
from __future__ import annotations
import argparse
import hashlib
import json
import math
import os
import subprocess
import sys
import tempfile
from collections.abc import Mapping, Sequence
from pathlib import Path
from typing import Any
try:
import pyarrow as pa
import pyarrow.parquet as pq
except ModuleNotFoundError: # pragma: no cover - concise CLI error in main()
pa = None # type: ignore[assignment]
pq = None # type: ignore[assignment]
EXPECTED_ROWS = 90_447
DEFAULT_SHARD_SIZE = 10_000
UPSTREAM_REPO_ID = "hotpotqa/hotpot_qa"
UPSTREAM_CONFIG = "distractor"
UPSTREAM_SPLIT = "train"
UPSTREAM_REVISION = "1908d6afbbead072334abe2965f91bd2709910ab"
PATCH_MANIFEST = Path(__file__).resolve().parents[1] / "SOURCE_PATCHES.json"
def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
package_root = Path(__file__).resolve().parents[1]
parser = argparse.ArgumentParser(
description="Build and validate all Bactrainus HotpotQA train-only Parquet views."
)
parser.add_argument(
"source",
type=Path,
nargs="?",
help=(
"optional complete official HotpotQA train JSON/JSONL export; "
"omitted downloads the pinned Hugging Face source"
),
)
parser.add_argument(
"--root",
type=Path,
default=package_root,
help=f"dataset checkout root (default: {package_root})",
)
parser.add_argument(
"--shard-size",
type=int,
default=DEFAULT_SHARD_SIZE,
help=f"rows per Parquet shard (default: {DEFAULT_SHARD_SIZE})",
)
return parser.parse_args(argv)
def view_builders() -> tuple[tuple[str, Any], ...]:
"""Construct every task view approved by the release contract."""
try:
from bactrainus.data import (
CotReaderViewBuilder,
DecomposedSentenceSelectorViewBuilder,
JointViewBuilder,
ParagraphSelectorViewBuilder,
QuestionDecomposerViewBuilder,
ReaderViewBuilder,
SentenceSelectorViewBuilder,
StructuredViewBuilder,
)
except ModuleNotFoundError as error:
raise RuntimeError(
"Install the clean Bactrainus package before building the dataset"
) from error
return (
("structured", StructuredViewBuilder()),
("reader-sft", ReaderViewBuilder()),
("cot-reader-sft", CotReaderViewBuilder()),
("paragraph-selector-sft", ParagraphSelectorViewBuilder()),
("question-decomposer-sft", QuestionDecomposerViewBuilder()),
("sentence-selector-sft", SentenceSelectorViewBuilder()),
(
"decomposed-sentence-selector-sft",
DecomposedSentenceSelectorViewBuilder(),
),
("joint-selector-reader-sft", JointViewBuilder()),
)
def _sha256_file(path: Path, chunk_size: int = 1024 * 1024) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(chunk_size), b""):
digest.update(chunk)
return digest.hexdigest()
def load_annotation_patches() -> dict[tuple[str, str, int], int | None]:
"""Load the reviewed repairs for invalid upstream sentence indices."""
payload = json.loads(PATCH_MANIFEST.read_text(encoding="utf-8"))
if payload.get("upstream_revision") != UPSTREAM_REVISION:
raise ValueError(
"SOURCE_PATCHES.json does not match the pinned upstream revision"
)
declared = payload.get("invalid_fact_count")
records = payload.get("patches")
if not isinstance(records, list) or declared != len(records):
raise ValueError("SOURCE_PATCHES.json has an invalid patch count")
patches: dict[tuple[str, str, int], int | None] = {}
for record in records:
if not isinstance(record, Mapping):
raise TypeError("every source patch must be an object")
key = (
str(record["source_id"]),
str(record["title"]),
int(record["invalid_sentence_index"]),
)
if key in patches:
raise ValueError(f"duplicate source patch: {key!r}")
action = record.get("action")
if action == "drop_redundant":
patches[key] = None
elif action == "replace":
replacement = record.get("replacement_sentence_index")
if isinstance(replacement, bool) or not isinstance(replacement, int):
raise ValueError(f"source patch has invalid replacement: {key!r}")
patches[key] = replacement
else:
raise ValueError(f"source patch has invalid action: {key!r}")
return patches
def apply_annotation_patches(
raw: dict[str, Any],
patches: Mapping[tuple[str, str, int], int | None],
) -> dict[str, Any]:
"""Apply only manifest-listed repairs while preserving fact order."""
source_id = raw.get("_id", raw.get("id"))
resolved: list[list[Any]] = []
for title, sentence_index in raw["supporting_facts"]:
key = (source_id, title, sentence_index)
if key not in patches:
resolved.append([title, sentence_index])
continue
replacement = patches[key]
if replacement is not None:
resolved.append([title, replacement])
raw["supporting_facts"] = resolved
return raw
def _hub_row_to_raw(
row: dict[str, Any],
patches: Mapping[tuple[str, str, int], int | None],
) -> dict[str, Any]:
"""Convert the official Hugging Face feature layout to HotpotQA JSON."""
context = row["context"]
facts = row["supporting_facts"]
titles = context["title"]
sentences = context["sentences"]
fact_titles = facts["title"]
sentence_ids = facts["sent_id"]
if len(titles) != len(sentences):
raise ValueError(
f"context columns are misaligned for source ID {row.get('id')!r}"
)
if len(fact_titles) != len(sentence_ids):
raise ValueError(
f"supporting-fact columns are misaligned for source ID {row.get('id')!r}"
)
return apply_annotation_patches(
{
"_id": row["id"],
"question": row["question"],
"answer": row["answer"],
"type": row["type"],
"level": row["level"],
"context": [
[title, values] for title, values in zip(titles, sentences, strict=True)
],
"supporting_facts": [
[title, sentence_id]
for title, sentence_id in zip(fact_titles, sentence_ids, strict=True)
],
},
patches,
)
def _load_local_records(path: Path) -> list[dict[str, Any]]:
text = path.read_text(encoding="utf-8")
if not text.strip():
raise ValueError(f"source dataset is empty: {path}")
if text.lstrip().startswith("["):
payload = json.loads(text)
if not isinstance(payload, list):
raise ValueError("JSON dataset root must be a list")
candidates = payload
else:
candidates = [json.loads(line) for line in text.splitlines() if line.strip()]
if any(not isinstance(record, dict) for record in candidates):
raise ValueError("every source record must be a JSON object")
return candidates
def load_source_examples(source: Path | None) -> tuple[tuple[Any, ...], dict[str, Any]]:
"""Load a complete local export or the immutable official Hub revision."""
try:
from bactrainus.data import parse_hotpot_examples
except ModuleNotFoundError as error:
raise RuntimeError(
"Install the clean Bactrainus package before building the dataset"
) from error
patches = load_annotation_patches()
if source is not None:
resolved = source.resolve()
if not resolved.is_file():
raise FileNotFoundError(resolved)
records = [
apply_annotation_patches(record, patches)
for record in _load_local_records(resolved)
]
examples = parse_hotpot_examples(records, split=UPSTREAM_SPLIT)
provenance = {
"mode": "official-local-export",
"filename": resolved.name,
"bytes": resolved.stat().st_size,
"sha256": _sha256_file(resolved),
}
return examples, provenance
try:
from datasets import load_dataset
except ModuleNotFoundError as error:
raise RuntimeError(
"Install the 'datasets' package to build directly from Hugging Face"
) from error
dataset = load_dataset(
UPSTREAM_REPO_ID,
UPSTREAM_CONFIG,
split=UPSTREAM_SPLIT,
revision=UPSTREAM_REVISION,
)
examples = parse_hotpot_examples(
(_hub_row_to_raw(row, patches) for row in dataset), split=UPSTREAM_SPLIT
)
provenance = {
"mode": "huggingface-datasets",
"repository": UPSTREAM_REPO_ID,
"revision": UPSTREAM_REVISION,
"config": UPSTREAM_CONFIG,
"split": UPSTREAM_SPLIT,
}
return examples, provenance
def write_source_manifest(
destination: Path,
provenance: dict[str, Any],
configs: Sequence[str],
) -> None:
"""Write machine-readable source identity and release coverage."""
payload = {
"schema_version": 1,
"upstream": {**provenance, "row_count": EXPECTED_ROWS},
"annotation_patches": {
"manifest": PATCH_MANIFEST.name,
"invalid_fact_count": len(load_annotation_patches()),
"sha256": _sha256_file(PATCH_MANIFEST),
},
"release": {
"row_count_per_config": EXPECTED_ROWS,
"identity_key": "source_id",
"configs": list(configs),
},
}
destination.write_text(
json.dumps(payload, indent=2, ensure_ascii=False) + "\n",
encoding="utf-8",
)
def write_view(
examples: Sequence[Any],
builder: Any,
destination: Path,
shard_size: int,
) -> int:
"""Write one deterministic view with a stable schema across shards."""
assert pa is not None and pq is not None
destination.mkdir(parents=True, exist_ok=False)
shard_count = math.ceil(len(examples) / shard_size)
reference_schema: Any | None = None
for shard_index, start in enumerate(range(0, len(examples), shard_size)):
stop = min(start + shard_size, len(examples))
rows = [builder.build(example).to_dict() for example in examples[start:stop]]
table = pa.Table.from_pylist(rows, schema=reference_schema)
if reference_schema is None:
reference_schema = table.schema
shard_name = f"train-{shard_index:05d}-of-{shard_count:05d}.parquet"
pq.write_table(
table,
destination / shard_name,
compression="zstd",
use_dictionary=True,
write_statistics=True,
)
return shard_count
def build_release(source: Path | None, root: Path, shard_size: int) -> None:
"""Build in staging, validate, and atomically install release artifacts."""
if shard_size <= 0:
raise ValueError("--shard-size must be positive")
root = root.resolve()
if not root.is_dir():
raise FileNotFoundError(root)
target_data = root / "data"
target_manifest = root / "CHECKSUMS.sha256"
target_source_manifest = root / "SOURCE_MANIFEST.json"
if (
target_data.exists()
or target_manifest.exists()
or target_source_manifest.exists()
):
raise FileExistsError(
"Refusing to overwrite existing release artifacts; use a clean checkout"
)
examples, provenance = load_source_examples(source)
if len(examples) != EXPECTED_ROWS:
raise ValueError(
f"source contains {len(examples):,} records; expected {EXPECTED_ROWS:,}"
)
validator = Path(__file__).with_name("validate_release.py").resolve()
with tempfile.TemporaryDirectory(
prefix=".bactrainus-build-", dir=root
) as temporary:
staging = Path(temporary)
builders = view_builders()
for config, builder in builders:
shards = write_view(
examples,
builder,
staging / "data" / config,
shard_size,
)
print(f"built {config}: {len(examples):,} rows in {shards} shard(s)")
subprocess.run(
[sys.executable, str(validator), "--root", str(staging)],
check=True,
)
write_source_manifest(
staging / "SOURCE_MANIFEST.json",
provenance,
[config for config, _ in builders],
)
os.replace(staging / "data", target_data)
os.replace(staging / "CHECKSUMS.sha256", target_manifest)
os.replace(staging / "SOURCE_MANIFEST.json", target_source_manifest)
def main(argv: Sequence[str] | None = None) -> int:
args = parse_args(argv)
if pa is None or pq is None:
print(
"error: pyarrow is required; install Bactrainus with the 'data' extra",
file=sys.stderr,
)
return 2
try:
build_release(args.source, args.root, args.shard_size)
except (FileNotFoundError, FileExistsError, RuntimeError, ValueError) as error:
print(f"error: {error}", file=sys.stderr)
return 1
print("Release build completed only after full validation.")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|