de-wiki / clean_wikipedia_jsonl.py
gmrrh's picture
Upload folder using huggingface_hub
12374cc verified
Raw History Blame Contribute Delete
10 kB
"""
clean_wikipedia_jsonl.py
========================
Post-process the JSONL output of fetch_wikipedia_de.py to fix common
artifacts left behind by `mwparserfromhell.strip_code()` and apply a
tighter length cap.
Cleanup patterns applied (defaults — all individually disable-able)
-------------------------------------------------------------------
1. Image-cell residue: lines starting with image-related prefixes
(`thumb|`, `mini|`, `rahmenlos|`, `zentriert|`, `links|`, `rechts|`,
`gerahmt|`, `rahmen|`, `border|`, `framed|`, `upright|`, ...)
followed by a pipe. Typical MediaWiki image-cell syntax.
2. Image-size-only cells: lines like `|300px|Description` that survive
strip_code as standalone table cells.
3. External-reference footnote lines: German Wikipedia references of
the form "Erklärvideo zu ... auf Youtube, abgerufen am 7. April
2024". Matched by the `abgerufen am` phrase with word boundaries.
4. Wiki category tags: `Kategorie:...` or `Category:...` lines.
5. File / image namespace links: `Datei:...`, `File:...`, `Bild:...`,
`Image:...` lines.
6. Whitespace runs: 3+ consecutive newlines collapsed to 2.
After cleanup:
7. Drop articles below --min-length or above --max-length chars.
8. Drop articles whose text became empty after cleanup.
Streams line-by-line: safe for 10+ GB inputs, peak RAM is one article
at a time.
Usage
-----
# Default cleanup (all patterns enabled, max-length 80k)
python clean_wikipedia_jsonl.py in.jsonl out.jsonl
# Soft cap (keep long articles)
python clean_wikipedia_jsonl.py in.jsonl out.jsonl --max-length 500000
# Disable specific cleanups (e.g. for diff comparisons)
python clean_wikipedia_jsonl.py in.jsonl out.jsonl --no-strip-image-cells
python clean_wikipedia_jsonl.py in.jsonl out.jsonl --no-strip-external-refs
python clean_wikipedia_jsonl.py in.jsonl out.jsonl --no-strip-categories
python clean_wikipedia_jsonl.py in.jsonl out.jsonl --no-strip-file-links
# Tighten both ends at once
python clean_wikipedia_jsonl.py in.jsonl out.jsonl --min-length 2000 \
--max-length 50000
Stdlib only.
"""
from __future__ import annotations
import argparse
import json
import re
import sys
import time
from pathlib import Path
# 1. Image-cell residue: lines starting with image-related prefixes
# followed by a pipe and the cell content.
IMAGE_CELL_PATTERN = re.compile(
r"^[ \t]*(?:thumb|mini|rahmenlos|rahmen|zentriert|gerahmt|"
r"border|framed|upright|links?|rechts?|vergrößern|verkleinern)"
r"\|[^\n]*$",
re.MULTILINE,
)
# 2. Image-size-only cell: "|300px|Description" or "300px|Description"
IMAGE_SIZE_PATTERN = re.compile(
r"^[ \t]*\|?\d+px\|[^\n]*$",
re.MULTILINE,
)
# 3. External-reference footnote: "abgerufen am <date>" (German Wikipedia)
EXTERNAL_REF_PATTERN = re.compile(
r"^[ \t]*[^\n]*\babgerufen am\b[^\n]*$",
re.MULTILINE,
)
# 4. Wiki category tag: "Kategorie:..." or "Category:..."
CATEGORY_PATTERN = re.compile(
r"^[ \t]*(?:Kategorie|Category):[^\n]*$",
re.MULTILINE,
)
# 5. File / image namespace link residue: "Datei:...", "File:...", etc.
FILE_LINK_PATTERN = re.compile(
r"^[ \t]*(?:Datei|File|Bild|Image|Dateiarchiv):[^\n]*$",
re.MULTILINE,
)
# 6. Whitespace runs: 3+ newlines -> 2
WHITESPACE_RUN = re.compile(r"\n{3,}")
def clean_text(text: str, *, strip_image_cells: bool,
strip_external_refs: bool, strip_categories: bool,
strip_file_links: bool) -> tuple[str, dict]:
"""Apply enabled cleanup patterns; return (cleaned_text, per-pattern_counts)."""
counts = {
"image_cells": 0,
"image_sizes": 0,
"external_refs": 0,
"categories": 0,
"file_links": 0,
}
if strip_image_cells:
text, n = IMAGE_CELL_PATTERN.subn("", text)
counts["image_cells"] = n
text, n = IMAGE_SIZE_PATTERN.subn("", text)
counts["image_sizes"] = n
if strip_external_refs:
text, n = EXTERNAL_REF_PATTERN.subn("", text)
counts["external_refs"] = n
if strip_categories:
text, n = CATEGORY_PATTERN.subn("", text)
counts["categories"] = n
if strip_file_links:
text, n = FILE_LINK_PATTERN.subn("", text)
counts["file_links"] = n
text = WHITESPACE_RUN.sub("\n\n", text)
return text.strip(), counts
def process(args: argparse.Namespace) -> None:
src = Path(args.input)
dst = Path(args.output)
if not src.is_file():
print(f"[error] Input not found: {src}", file=sys.stderr)
sys.exit(1)
dst.parent.mkdir(parents=True, exist_ok=True)
print(f"[info] {src} ({src.stat().st_size:,} bytes)", file=sys.stderr)
print(f"[info] -> {dst}", file=sys.stderr)
print(
f"[info] min_length={args.min_length} max_length={args.max_length}",
file=sys.stderr,
)
print(
f"[info] image_cells={not args.no_strip_image_cells} "
f"external_refs={not args.no_strip_external_refs} "
f"categories={not args.no_strip_categories} "
f"file_links={not args.no_strip_file_links}",
file=sys.stderr,
)
stats = {
"seen": 0, "kept": 0,
"drop_invalid": 0, "drop_empty": 0,
"drop_too_short": 0, "drop_too_long": 0,
"lines_stripped": {
"image_cells": 0,
"image_sizes": 0,
"external_refs": 0,
"categories": 0,
"file_links": 0,
},
"chars_before": 0,
"chars_after": 0,
}
started = time.time()
with open(src, "rb") as fin, open(dst, "w", encoding="utf-8") as fout:
for raw in fin:
stats["seen"] += 1
try:
ex = json.loads(raw)
except json.JSONDecodeError:
stats["drop_invalid"] += 1
continue
if not isinstance(ex, dict) or "text" not in ex:
stats["drop_invalid"] += 1
continue
original = ex["text"]
if not isinstance(original, str):
stats["drop_invalid"] += 1
continue
stats["chars_before"] += len(original)
cleaned, counts = clean_text(
original,
strip_image_cells=not args.no_strip_image_cells,
strip_external_refs=not args.no_strip_external_refs,
strip_categories=not args.no_strip_categories,
strip_file_links=not args.no_strip_file_links,
)
for k, v in counts.items():
stats["lines_stripped"][k] += v
if not cleaned:
stats["drop_empty"] += 1
continue
if len(cleaned) < args.min_length:
stats["drop_too_short"] += 1
continue
if len(cleaned) > args.max_length:
stats["drop_too_long"] += 1
continue
stats["chars_after"] += len(cleaned)
fout.write(json.dumps({"text": cleaned}, ensure_ascii=False) + "\n")
stats["kept"] += 1
if stats["seen"] % 500_000 == 0:
elapsed = time.time() - started
rate = stats["seen"] / elapsed if elapsed > 0 else 0.0
print(
f"[progress] {stats['seen']:>10} seen | "
f"{stats['kept']:>10} kept | "
f"{rate:>7,.0f} lines/s",
file=sys.stderr,
)
elapsed = time.time() - started
rate = stats["seen"] / elapsed if elapsed > 0 else 0.0
pct_kept = 100 * stats["kept"] / max(stats["seen"], 1)
pct_chars = (
100 * stats["chars_after"] / stats["chars_before"]
if stats["chars_before"] else 0.0
)
print("\n[done] " + "=" * 56, file=sys.stderr)
print(f" Articles seen : {stats['seen']:,}", file=sys.stderr)
print(f" Articles kept : {stats['kept']:,} ({pct_kept:.2f}%)",
file=sys.stderr)
print(f" Chars in -> out : "
f"{stats['chars_before']:,} -> {stats['chars_after']:,} "
f"({pct_chars:.2f}%)",
file=sys.stderr)
print(f" Dropped (invalid): {stats['drop_invalid']:,}", file=sys.stderr)
print(f" Dropped (empty) : {stats['drop_empty']:,}", file=sys.stderr)
print(f" Dropped (short) : {stats['drop_too_short']:,}", file=sys.stderr)
print(f" Dropped (long) : {stats['drop_too_long']:,}", file=sys.stderr)
print(f" Lines stripped :", file=sys.stderr)
for k, v in stats["lines_stripped"].items():
print(f" {k:<14}: {v:,}", file=sys.stderr)
print(f" Elapsed : {elapsed:.1f}s ({rate:,.0f} lines/s)",
file=sys.stderr)
print("=" * 56, file=sys.stderr)
def main() -> None:
p = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter,
)
p.add_argument("input", help="Input JSONL file (e.g. wikipedia_de_clean.jsonl)")
p.add_argument("output", help="Output JSONL file (overwritten)")
p.add_argument("--max-length", type=int, default=80_000,
help="Drop articles above N chars (default: 80000)")
p.add_argument("--min-length", type=int, default=1_500,
help="Drop articles below N chars (default: 1500)")
p.add_argument("--no-strip-image-cells", action="store_true",
help="Disable image-cell residue stripping (default: enabled)")
p.add_argument("--no-strip-external-refs", action="store_true",
help="Disable external-reference stripping (default: enabled)")
p.add_argument("--no-strip-categories", action="store_true",
help="Disable category-line stripping (default: enabled)")
p.add_argument("--no-strip-file-links", action="store_true",
help="Disable file/image-namespace link stripping (default: enabled)")
args = p.parse_args()
process(args)
if __name__ == "__main__":
main()