Datasets:
Tasks:
Text Generation
Formats:
parquet
Sub-tasks:
language-modeling
Languages:
German
Size:
100M - 1B
ArXiv:
Tags:
text-corpus
continual-development
community-collaboration
synthetic-annotations
work-in-progress
License:
Commit ·
4ca878d
1
Parent(s): f34e983
cellar: parse in a forked child; gutenberg: licence section as H2
Browse filesBoth converters are Rust and abort the worker on some files, which killed the
map with 'subprocess abruptly died' every few minutes; forking per document
turns that into a skipped document, with a timeout for the ones that parse for
minutes. The Gutenberg card's licence block sat under an H3, so the datasheet
test that requires licence_information for license: other could not find it.
- data/cellar/create.py +367 -0
- data/gutenberg/gutenberg.md +139 -0
data/cellar/create.py
ADDED
|
@@ -0,0 +1,367 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# /// script
|
| 2 |
+
# requires-python = ">=3.12"
|
| 3 |
+
# dependencies = [
|
| 4 |
+
# "datasets>=3.0.0",
|
| 5 |
+
# "requests",
|
| 6 |
+
# "tenacity",
|
| 7 |
+
# "multiprocess",
|
| 8 |
+
# "liteparse",
|
| 9 |
+
# "html-to-markdown",
|
| 10 |
+
# "pyarrow>=17.0",
|
| 11 |
+
# "dynaword>=0.1",
|
| 12 |
+
# ]
|
| 13 |
+
# ///
|
| 14 |
+
"""Fetch the German documents of the EU Publications Office (Cellar).
|
| 15 |
+
|
| 16 |
+
To run this script from the repository root:
|
| 17 |
+
|
| 18 |
+
uv run data/cellar/create.py
|
| 19 |
+
|
| 20 |
+
The pipeline is the one `danish-dynaword` and `swedish-dynaword` use for the same
|
| 21 |
+
collection, with the language code switched to `DEU`:
|
| 22 |
+
|
| 23 |
+
- the document list comes from `danish-foundation-models/cellar-metadata`, which
|
| 24 |
+
holds one row per (document, language, file type) with the download URL and the
|
| 25 |
+
document date;
|
| 26 |
+
- PDFs are converted with LiteParse (no OCR, 100-page cap) and markup files with
|
| 27 |
+
`html-to-markdown`; anything else is skipped;
|
| 28 |
+
- an item published in several manifestations is kept once, in the format that
|
| 29 |
+
sorts first — the same rule the Danish subset applies, so an act is one document
|
| 30 |
+
rather than one document per file format.
|
| 31 |
+
|
| 32 |
+
Cellar is published under CC-BY-SA 4.0; every document here is licensed that way,
|
| 33 |
+
which is why the subset card says so and the `license` column repeats it per row.
|
| 34 |
+
"""
|
| 35 |
+
|
| 36 |
+
from __future__ import annotations
|
| 37 |
+
|
| 38 |
+
import logging
|
| 39 |
+
import os
|
| 40 |
+
import re
|
| 41 |
+
import select
|
| 42 |
+
import signal
|
| 43 |
+
import time
|
| 44 |
+
from datetime import date
|
| 45 |
+
from pathlib import Path
|
| 46 |
+
|
| 47 |
+
import multiprocess
|
| 48 |
+
import pyarrow as pa
|
| 49 |
+
import pyarrow.parquet as pq
|
| 50 |
+
import requests
|
| 51 |
+
from datasets import Dataset, load_dataset, load_from_disk
|
| 52 |
+
from html_to_markdown import ConversionOptions, convert
|
| 53 |
+
from liteparse import LiteParse
|
| 54 |
+
from tenacity import retry, stop_after_attempt, wait_exponential
|
| 55 |
+
|
| 56 |
+
from dynaword.dataset_structure import COLUMN_ORDER
|
| 57 |
+
from dynaword.process_dataset import (
|
| 58 |
+
add_token_count,
|
| 59 |
+
remove_duplicate_text,
|
| 60 |
+
remove_empty_texts,
|
| 61 |
+
)
|
| 62 |
+
|
| 63 |
+
logger = logging.getLogger(__name__)
|
| 64 |
+
|
| 65 |
+
SOURCE = "cellar"
|
| 66 |
+
WORKERS = int(os.environ.get("CELLAR_WORKERS", 32))
|
| 67 |
+
LANGCODE = "DEU"
|
| 68 |
+
LICENSE = "CC-BY-SA-4.0"
|
| 69 |
+
METADATA_REPO = "danish-foundation-models/cellar-metadata"
|
| 70 |
+
METADATA_REVISION = "0601edf401679f30079c4df7fcf752602b4498fb"
|
| 71 |
+
# documents without a date in the metadata get the collection window, as in the
|
| 72 |
+
# Danish and Swedish subsets
|
| 73 |
+
FALLBACK_CREATED = "1950-01-01, 2026-12-31"
|
| 74 |
+
PDF_TYPES = {"pdf", "pdfa2a", "pdfa1a", "pdf1x", "pdfx4", "pdfx", "pdfa1b"}
|
| 75 |
+
MARKUP_TYPES = {"html", "xhtml", "xhtml5", "xml", "etsi_xml"}
|
| 76 |
+
# C0/C1 control characters and U+FFFD: what a binary blob turns into after a
|
| 77 |
+
# "successful" conversion
|
| 78 |
+
NON_TEXT = re.compile(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f-\x9f\ufffd]")
|
| 79 |
+
|
| 80 |
+
ANNOTATION_COLUMNS = [
|
| 81 |
+
"content_integrity",
|
| 82 |
+
"content_ratio",
|
| 83 |
+
"content_length",
|
| 84 |
+
"one_sentence_description",
|
| 85 |
+
"content_type",
|
| 86 |
+
"business_sector",
|
| 87 |
+
"technical_content",
|
| 88 |
+
"information_density",
|
| 89 |
+
"content_quality",
|
| 90 |
+
"audience_level",
|
| 91 |
+
"commercial_bias",
|
| 92 |
+
"time_sensitivity",
|
| 93 |
+
"content_safety",
|
| 94 |
+
"educational_value",
|
| 95 |
+
"reasoning_indicators",
|
| 96 |
+
"pii_presence",
|
| 97 |
+
"regional_relevance",
|
| 98 |
+
"country_relevance",
|
| 99 |
+
]
|
| 100 |
+
|
| 101 |
+
ROOT = Path(__file__).parent
|
| 102 |
+
CACHE = Path.home() / ".cache" / "dynaword" / SOURCE / METADATA_REVISION
|
| 103 |
+
EXTRACT_OUT = CACHE / "extracted"
|
| 104 |
+
DS_PATH = CACHE / f"german_metadata"
|
| 105 |
+
|
| 106 |
+
CRASHED_FILE = CACHE / "crashed.txt"
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
def load_crashed() -> set[int]:
|
| 110 |
+
"""The indices whose conversion killed the worker process.
|
| 111 |
+
|
| 112 |
+
A parser abort inside liteparse (a Rust panic on a malformed file) kills the
|
| 113 |
+
worker without raising, so datasets only reports that a subprocess died. The
|
| 114 |
+
per-document cache makes a restart cheap, but the map reaches the same
|
| 115 |
+
document again and dies at the same place, so the build never gets past it.
|
| 116 |
+
Every parse attempt leaves an .inprogress marker; a marker without its .md
|
| 117 |
+
names a document that killed a worker. Record it once and skip it, so a
|
| 118 |
+
restart makes progress instead of repeating the same death.
|
| 119 |
+
"""
|
| 120 |
+
crashed: set[int] = set()
|
| 121 |
+
if CRASHED_FILE.exists():
|
| 122 |
+
crashed |= {
|
| 123 |
+
int(line) for line in CRASHED_FILE.read_text().split() if line.strip()
|
| 124 |
+
}
|
| 125 |
+
for marker in sorted(EXTRACT_OUT.glob("*.inprogress")):
|
| 126 |
+
index = int(marker.stem)
|
| 127 |
+
if not (EXTRACT_OUT / f"{index}.md").exists():
|
| 128 |
+
crashed.add(index)
|
| 129 |
+
with CRASHED_FILE.open("a") as handle:
|
| 130 |
+
handle.write(f"{index}\n")
|
| 131 |
+
marker.unlink()
|
| 132 |
+
if crashed:
|
| 133 |
+
logger.info("cellar: %s documents known to abort the parser", len(crashed))
|
| 134 |
+
return crashed
|
| 135 |
+
|
| 136 |
+
|
| 137 |
+
CRASHED = load_crashed()
|
| 138 |
+
|
| 139 |
+
CONVERT_TIMEOUT = float(os.environ.get("CELLAR_CONVERT_TIMEOUT", 180))
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
def convert_isolated(parse, timeout: float = CONVERT_TIMEOUT) -> str | None:
|
| 143 |
+
"""Run a parser in a forked child and never let it take the worker down.
|
| 144 |
+
|
| 145 |
+
Both converters are Rust extensions, and on a malformed file one of them
|
| 146 |
+
kills the process outright instead of raising: the dataset map then dies
|
| 147 |
+
with "one of the subprocesses has abruptly died", minutes of work are lost,
|
| 148 |
+
and the restart walks into the same file. Some files also parse for many
|
| 149 |
+
minutes. A fork per document turns both into a skipped document with a log
|
| 150 |
+
line, which is all a corpus wants from a file it cannot read.
|
| 151 |
+
"""
|
| 152 |
+
read_fd, write_fd = os.pipe()
|
| 153 |
+
pid = os.fork()
|
| 154 |
+
if pid == 0:
|
| 155 |
+
os.close(read_fd)
|
| 156 |
+
try:
|
| 157 |
+
payload = (parse() or "").encode("utf-8", "replace")
|
| 158 |
+
except BaseException as error:
|
| 159 |
+
payload = b"\0" + str(error).encode("utf-8", "replace")[:500]
|
| 160 |
+
try:
|
| 161 |
+
with os.fdopen(write_fd, "wb") as handle:
|
| 162 |
+
handle.write(payload)
|
| 163 |
+
except BaseException:
|
| 164 |
+
os._exit(1)
|
| 165 |
+
os._exit(0)
|
| 166 |
+
os.close(write_fd)
|
| 167 |
+
chunks: list[bytes] = []
|
| 168 |
+
deadline = time.monotonic() + timeout
|
| 169 |
+
try:
|
| 170 |
+
while True:
|
| 171 |
+
remaining = deadline - time.monotonic()
|
| 172 |
+
if remaining <= 0:
|
| 173 |
+
os.kill(pid, signal.SIGKILL)
|
| 174 |
+
print("Parse timed out, skipping...")
|
| 175 |
+
return None
|
| 176 |
+
ready, _, _ = select.select([read_fd], [], [], min(remaining, 5.0))
|
| 177 |
+
if ready:
|
| 178 |
+
block = os.read(read_fd, 1 << 20)
|
| 179 |
+
if not block:
|
| 180 |
+
break
|
| 181 |
+
chunks.append(block)
|
| 182 |
+
finally:
|
| 183 |
+
os.close(read_fd)
|
| 184 |
+
_, status = os.waitpid(pid, 0)
|
| 185 |
+
if not (os.WIFEXITED(status) and os.WEXITSTATUS(status) == 0):
|
| 186 |
+
print("Parser aborted, skipping...")
|
| 187 |
+
return None
|
| 188 |
+
payload = b"".join(chunks)
|
| 189 |
+
if payload.startswith(b"\0"):
|
| 190 |
+
raise RuntimeError(payload[1:].decode("utf-8", "replace"))
|
| 191 |
+
return payload.decode("utf-8", "replace")
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
@retry(stop=stop_after_attempt(3), wait=wait_exponential(multiplier=1))
|
| 195 |
+
def fetch(url: str):
|
| 196 |
+
with requests.get(url, timeout=30, stream=True) as response:
|
| 197 |
+
response.raise_for_status()
|
| 198 |
+
limit = 100 * 1024 * 1024
|
| 199 |
+
if int(response.headers.get("Content-Length", 0)) > limit:
|
| 200 |
+
return None
|
| 201 |
+
content = bytearray()
|
| 202 |
+
for block in response.iter_content(1024 * 1024):
|
| 203 |
+
content.extend(block)
|
| 204 |
+
if len(content) > limit:
|
| 205 |
+
return None
|
| 206 |
+
response._content = bytes(content)
|
| 207 |
+
return response
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
def parse_uri(uri: str) -> tuple[str, int]:
|
| 211 |
+
"""The item id and manifestation number of a Cellar URI."""
|
| 212 |
+
item = uri.rsplit("/", 2)[1]
|
| 213 |
+
item_id, _expression, manifestation = item.rsplit(".", 2)
|
| 214 |
+
return item_id, int(manifestation)
|
| 215 |
+
|
| 216 |
+
|
| 217 |
+
def fetch_and_convert(example: dict) -> None:
|
| 218 |
+
file_type = example["type"]
|
| 219 |
+
url = example["item"]
|
| 220 |
+
index = example["global_index"]
|
| 221 |
+
|
| 222 |
+
if file_type not in PDF_TYPES | MARKUP_TYPES:
|
| 223 |
+
return
|
| 224 |
+
output = EXTRACT_OUT / f"{index}.md"
|
| 225 |
+
if output.exists() and output.stat().st_size:
|
| 226 |
+
return
|
| 227 |
+
if index in CRASHED:
|
| 228 |
+
return
|
| 229 |
+
mark = EXTRACT_OUT / f"{index}.inprogress"
|
| 230 |
+
|
| 231 |
+
try:
|
| 232 |
+
response = fetch(url)
|
| 233 |
+
if response is None:
|
| 234 |
+
return
|
| 235 |
+
except Exception as error: # noqa: BLE001 - one bad document must not stop the run
|
| 236 |
+
print(f"Fetch exception, skipping... {error}")
|
| 237 |
+
return
|
| 238 |
+
|
| 239 |
+
mark.write_text("")
|
| 240 |
+
try:
|
| 241 |
+
if file_type in PDF_TYPES:
|
| 242 |
+
text = convert_isolated(
|
| 243 |
+
lambda: LiteParse(
|
| 244 |
+
output_format="markdown",
|
| 245 |
+
image_mode="off",
|
| 246 |
+
extract_links=False,
|
| 247 |
+
quiet=True,
|
| 248 |
+
max_pages=100,
|
| 249 |
+
ocr_enabled=False,
|
| 250 |
+
).parse(response.content).text
|
| 251 |
+
)
|
| 252 |
+
else:
|
| 253 |
+
text = convert_isolated(
|
| 254 |
+
lambda: convert(
|
| 255 |
+
response.text,
|
| 256 |
+
ConversionOptions(extract_metadata=False, skip_images=True),
|
| 257 |
+
).content
|
| 258 |
+
)
|
| 259 |
+
if text and text.strip():
|
| 260 |
+
temporary = output.with_suffix(".tmp")
|
| 261 |
+
temporary.write_text(text, encoding="utf-8")
|
| 262 |
+
temporary.replace(output)
|
| 263 |
+
except Exception as error: # noqa: BLE001
|
| 264 |
+
print(f"Failed parsing {file_type} {url}: {error}")
|
| 265 |
+
finally:
|
| 266 |
+
mark.unlink(missing_ok=True)
|
| 267 |
+
|
| 268 |
+
|
| 269 |
+
def looks_like_text(text: str) -> bool:
|
| 270 |
+
"""Reject the rows whose URL serves a scan or a JPEG instead of the document.
|
| 271 |
+
|
| 272 |
+
Cellar's metadata declares a file type, but a few items are served as binary
|
| 273 |
+
blobs (the markup converter happily turns a JPEG into 1.7M characters of
|
| 274 |
+
noise). Anything with more than 2% non-text characters, or shorter than a
|
| 275 |
+
short paragraph, is dropped.
|
| 276 |
+
"""
|
| 277 |
+
sample = text[:200_000]
|
| 278 |
+
if len(sample) < 200:
|
| 279 |
+
return False
|
| 280 |
+
non_text = len(NON_TEXT.findall(sample))
|
| 281 |
+
return non_text / len(sample) < 0.02
|
| 282 |
+
|
| 283 |
+
|
| 284 |
+
def clean_samples(example: dict) -> dict:
|
| 285 |
+
example["text"] = re.sub(r"<\?xml.*?\?>", "", example["text"]).strip()
|
| 286 |
+
return example
|
| 287 |
+
|
| 288 |
+
|
| 289 |
+
def create_samples(metadata: Dataset):
|
| 290 |
+
date_map = dict(zip(metadata["global_index"], metadata["date"]))
|
| 291 |
+
item_map = dict(zip(metadata["global_index"], metadata["item"]))
|
| 292 |
+
scanned_items: dict[str, int] = {}
|
| 293 |
+
|
| 294 |
+
for file in sorted(EXTRACT_OUT.glob("*.md"), key=lambda path: int(path.stem)):
|
| 295 |
+
index = int(file.stem)
|
| 296 |
+
if index not in item_map:
|
| 297 |
+
continue
|
| 298 |
+
document = clean_samples({"text": file.read_text(encoding="utf-8")})["text"]
|
| 299 |
+
if not document or not looks_like_text(document):
|
| 300 |
+
continue
|
| 301 |
+
if document.count("\ufffd") > len(document) * 0.2:
|
| 302 |
+
continue
|
| 303 |
+
item_id, manifestation = parse_uri(item_map[index])
|
| 304 |
+
if scanned_items.setdefault(item_id, manifestation) != manifestation:
|
| 305 |
+
continue
|
| 306 |
+
written = date_map.get(index)
|
| 307 |
+
yield {
|
| 308 |
+
"id": f"{SOURCE}_{index}",
|
| 309 |
+
"text": document,
|
| 310 |
+
"source": SOURCE,
|
| 311 |
+
"license": LICENSE,
|
| 312 |
+
"added": date.today().isoformat(),
|
| 313 |
+
"created": f"{written}, {written}" if written else FALLBACK_CREATED,
|
| 314 |
+
}
|
| 315 |
+
|
| 316 |
+
|
| 317 |
+
def write_metadata(ids: list[str], path: Path) -> None:
|
| 318 |
+
table = pa.table(
|
| 319 |
+
{
|
| 320 |
+
"dataset": pa.array([SOURCE] * len(ids), type=pa.string()),
|
| 321 |
+
"id": pa.array(ids, type=pa.string()),
|
| 322 |
+
**{
|
| 323 |
+
column: pa.array([None] * len(ids), type=pa.string())
|
| 324 |
+
for column in ANNOTATION_COLUMNS
|
| 325 |
+
},
|
| 326 |
+
}
|
| 327 |
+
)
|
| 328 |
+
pq.write_table(table, path)
|
| 329 |
+
|
| 330 |
+
|
| 331 |
+
def main() -> None:
|
| 332 |
+
EXTRACT_OUT.mkdir(parents=True, exist_ok=True)
|
| 333 |
+
if DS_PATH.exists():
|
| 334 |
+
filtered = load_from_disk(DS_PATH)
|
| 335 |
+
else:
|
| 336 |
+
dataset = load_dataset(
|
| 337 |
+
METADATA_REPO, revision=METADATA_REVISION, split="train"
|
| 338 |
+
)
|
| 339 |
+
filtered = dataset.filter(lambda row: row["langCode"] == LANGCODE)
|
| 340 |
+
filtered = filtered.map(
|
| 341 |
+
lambda example, index: {"global_index": index}, with_indices=True
|
| 342 |
+
)
|
| 343 |
+
filtered.save_to_disk(DS_PATH)
|
| 344 |
+
logger.info("%s: %s documents in the metadata", SOURCE, len(filtered))
|
| 345 |
+
|
| 346 |
+
multiprocess.set_start_method("spawn", force=True)
|
| 347 |
+
filtered.map(
|
| 348 |
+
fetch_and_convert, desc="Fetching and converting...", num_proc=WORKERS
|
| 349 |
+
)
|
| 350 |
+
|
| 351 |
+
dataset = Dataset.from_generator(create_samples, gen_kwargs={"metadata": filtered})
|
| 352 |
+
dataset = remove_empty_texts(dataset)
|
| 353 |
+
dataset = remove_duplicate_text(dataset)
|
| 354 |
+
dataset = add_token_count(dataset)
|
| 355 |
+
dataset = dataset.select_columns([*COLUMN_ORDER, "license"])
|
| 356 |
+
dataset.to_parquet(ROOT / "data.parquet")
|
| 357 |
+
write_metadata(dataset["id"], ROOT / "metadata.parquet")
|
| 358 |
+
logger.info("%s: wrote %s documents", SOURCE, len(dataset))
|
| 359 |
+
|
| 360 |
+
|
| 361 |
+
if __name__ == "__main__":
|
| 362 |
+
started = time.strftime("%Y-%m-%d %H:%M:%S")
|
| 363 |
+
logging.basicConfig(
|
| 364 |
+
level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
|
| 365 |
+
)
|
| 366 |
+
logger.info("cellar build started %s", started)
|
| 367 |
+
main()
|
data/gutenberg/gutenberg.md
ADDED
|
@@ -0,0 +1,139 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
pretty_name: Project Gutenberg (German)
|
| 3 |
+
language:
|
| 4 |
+
- de
|
| 5 |
+
license: other
|
| 6 |
+
license_name: Gutenberg
|
| 7 |
+
size_categories:
|
| 8 |
+
- 1K<n<10K
|
| 9 |
+
task_categories:
|
| 10 |
+
- text-generation
|
| 11 |
+
- fill-mask
|
| 12 |
+
task_ids:
|
| 13 |
+
- language-modeling
|
| 14 |
+
source_datasets:
|
| 15 |
+
- original
|
| 16 |
+
domains:
|
| 17 |
+
- Books
|
| 18 |
+
---
|
| 19 |
+
|
| 20 |
+
# Dataset Card for Project Gutenberg (German)
|
| 21 |
+
|
| 22 |
+
<!-- START-SHORT DESCRIPTION -->
|
| 23 |
+
German-language books from [Project Gutenberg](https://www.gutenberg.org) whose authors died in 1955 or earlier
|
| 24 |
+
<!-- END-SHORT DESCRIPTION -->
|
| 25 |
+
|
| 26 |
+
Project Gutenberg is the oldest digital library of public-domain books. This
|
| 27 |
+
subset holds its German-language titles: prose, drama, poetry and non-fiction,
|
| 28 |
+
mostly 19th century, including the canonical authors (Goethe, Schiller, Kleist,
|
| 29 |
+
Rilke, Fontane) and many titles that are otherwise only available as scans.
|
| 30 |
+
|
| 31 |
+
## Dataset Description
|
| 32 |
+
|
| 33 |
+
<!-- START-DESC-STATS -->
|
| 34 |
+
- **Number of samples**: 1.78K
|
| 35 |
+
- **Number of tokens (Llama 3)**: 198.60M
|
| 36 |
+
- **Average document length in tokens (min, max)**: 111.26K (69, 1.19M)
|
| 37 |
+
<!-- END-DESC-STATS -->
|
| 38 |
+
|
| 39 |
+
## Dataset Structure
|
| 40 |
+
|
| 41 |
+
An example from the dataset looks as follows.
|
| 42 |
+
|
| 43 |
+
<!-- START-SAMPLE -->
|
| 44 |
+
```py
|
| 45 |
+
{
|
| 46 |
+
"id": "gutenberg_2054",
|
| 47 |
+
"text": "Iphigenie auf Tauris.\n\nEin Schauspiel.\n\nJohann Wolfgang von Goethe\n\nPersonen:\n\nIphigenie.\nThoas, Kön[...]",
|
| 48 |
+
"source": "gutenberg",
|
| 49 |
+
"added": "2026-10-30",
|
| 50 |
+
"created": "1749-01-01, 1832-12-31",
|
| 51 |
+
"token_count": 30092,
|
| 52 |
+
"license": "public-domain"
|
| 53 |
+
}
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
### Data Fields
|
| 57 |
+
|
| 58 |
+
An entry in the dataset consists of the following fields:
|
| 59 |
+
|
| 60 |
+
- `id` (`str`): A unique identifier for each document.
|
| 61 |
+
- `text` (`str`): The content of the document.
|
| 62 |
+
- `source` (`str`): The source of the document (see [Source Data](#source-data)).
|
| 63 |
+
- `added` (`str`): The date when the document was added to this collection.
|
| 64 |
+
- `created` (`str`): The date range when the document was originally created.
|
| 65 |
+
- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
|
| 66 |
+
<!-- END-SAMPLE -->
|
| 67 |
+
|
| 68 |
+
### Data Fields
|
| 69 |
+
|
| 70 |
+
An entry in the dataset consists of the following fields:
|
| 71 |
+
|
| 72 |
+
- `id` (`str`): A unique identifier for each document.
|
| 73 |
+
- `text` (`str`): The content of the document.
|
| 74 |
+
- `source` (`str`): The source of the document.
|
| 75 |
+
- `license` (`str`): the licence of the document, `public-domain` for this subset.
|
| 76 |
+
- `added` (`str`): The date when the document was added to this collection.
|
| 77 |
+
- `created` (`str`): The date range when the document was originally created.
|
| 78 |
+
- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
|
| 79 |
+
|
| 80 |
+
### Dataset Statistics
|
| 81 |
+
|
| 82 |
+
<!-- START-DATASET PLOTS -->
|
| 83 |
+
<p align="center">
|
| 84 |
+
<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
|
| 85 |
+
</p>
|
| 86 |
+
<!-- END-DATASET PLOTS -->
|
| 87 |
+
|
| 88 |
+
### Annotation Overview
|
| 89 |
+
|
| 90 |
+
<!-- START-ANNOTATION PLOTS -->
|
| 91 |
+
<!-- END-ANNOTATION PLOTS -->
|
| 92 |
+
|
| 93 |
+
## License Information
|
| 94 |
+
|
| 95 |
+
The books are in the public domain and Project Gutenberg publishes them under the
|
| 96 |
+
[Project Gutenberg License](https://www.gutenberg.org/policy/license.html), which
|
| 97 |
+
permits redistribution of the texts. This subset strips Gutenberg's header,
|
| 98 |
+
transcriber credit and licence block from every file — which is what those terms
|
| 99 |
+
ask of a redistributor — and keeps the work's own text alone.
|
| 100 |
+
|
| 101 |
+
Beyond the US public-domain determination Gutenberg applies, the build enforces
|
| 102 |
+
the German term as well: only books whose authors died in 1955 or earlier are
|
| 103 |
+
included, so copyright has expired under the 70-years-*post-mortem-auctoris*
|
| 104 |
+
rule too.
|
| 105 |
+
|
| 106 |
+
## Additional Information
|
| 107 |
+
|
| 108 |
+
### Source Data
|
| 109 |
+
|
| 110 |
+
The titles and their bibliographic metadata come from Gutenberg's catalogue
|
| 111 |
+
dump (`pg_catalog.csv`), filtered to `Language == de` and `Type == Text`. Each
|
| 112 |
+
book is downloaded as its UTF-8 plain-text file (`ebooks/<id>.txt.utf-8`) and
|
| 113 |
+
Gutenberg's own boilerplate — header, transcriber credit and licence — is
|
| 114 |
+
removed between its `START OF THE PROJECT GUTENBERG EBOOK` and `END OF …`
|
| 115 |
+
markers, which is also what its licence asks of a redistributor. A file without
|
| 116 |
+
both markers is refused rather than shipped with the licence text inside it.
|
| 117 |
+
|
| 118 |
+
### Annotations
|
| 119 |
+
|
| 120 |
+
Not available: the propella annotation release does not cover Project Gutenberg,
|
| 121 |
+
so `metadata.parquet` carries nulls and every document keeps its text. The subset
|
| 122 |
+
can be annotated from its text in a later batch run.
|
| 123 |
+
|
| 124 |
+
### Limitations
|
| 125 |
+
|
| 126 |
+
- **The `created` range is a bound, not a publication date.** Gutenberg does not
|
| 127 |
+
publish composition or edition dates (the dates it does carry are transcription
|
| 128 |
+
dates), so the range written for each document is the author's life span —
|
| 129 |
+
the interval in which the text must have been written. It is deliberately wider
|
| 130 |
+
than the true publication date.
|
| 131 |
+
- **Only books whose author died by 1955 are included**, so that German copyright
|
| 132 |
+
(70 years *post mortem auctoris*) has expired as well as the US term Gutenberg
|
| 133 |
+
works from. Books whose authors carry no dates are excluded entirely: the term
|
| 134 |
+
cannot be asserted for them. Of 2,415 German text entries in the catalogue,
|
| 135 |
+
1,8xx survive that filter.
|
| 136 |
+
- The books are volunteer transcriptions, so the error rate varies by book;
|
| 137 |
+
`remove_duplicate_text` has already dropped duplicates created by re-releases.
|
| 138 |
+
- Translations are included when their translator's dates are known and expired,
|
| 139 |
+
so a small share of the text is translated rather than original German.
|