KeremTT's picture
Update code
c02657c verified
Raw History Blame Contribute Delete
3.25 kB
"""Text normalisation, hashing, PII scrubbing and cheap quality heuristics."""
import hashlib
import re
import unicodedata
_WS = re.compile(r"\s+")
_WORD = re.compile(r"\w+", re.UNICODE)
EMAIL = re.compile(r"\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}\b")
IPV4 = re.compile(r"\b(?:(?:25[0-5]|2[0-4]\d|1?\d?\d)\.){3}(?:25[0-5]|2[0-4]\d|1?\d?\d)\b")
PHONE = re.compile(r"(?<!\w)(?:\+?\d{1,3}[ .-]?)?(?:\(\d{3}\)|\d{3})[ .-]\d{3}[ .-]\d{2,4}(?!\w)")
TR_CHARS = set("çğıöşüÇĞİÖŞÜ")
TR_WORDS = {"ve", "bir", "bu", "için", "ile", "da", "de", "olarak", "çok", "daha", "gibi", "olan", "ne", "ama"}
EN_WORDS = {"the", "and", "of", "to", "is", "in", "that", "it", "for", "with", "as", "was", "on", "are"}
def normalize(text: str) -> str:
text = unicodedata.normalize("NFKC", text)
return _WS.sub(" ", text).strip()
def key(text: str) -> str:
"""Hash used for exact de-duplication (case/whitespace/punctuation-insensitive)."""
canon = " ".join(_WORD.findall(normalize(text).casefold()))
return hashlib.blake2b(canon.encode(), digest_size=16).hexdigest()
def stable_id(*parts: str) -> str:
return hashlib.sha256("\x1f".join(parts).encode()).hexdigest()[:24]
def words(text: str):
return _WORD.findall(text.casefold())
def ngrams(tokens, n):
return {" ".join(tokens[i:i + n]) for i in range(len(tokens) - n + 1)}
def scrub_pii(text: str) -> str:
text = EMAIL.sub("<EMAIL>", text)
text = IPV4.sub("<IP_ADDRESS>", text)
return text
def guess_language(text: str) -> str:
sample = text[:4000]
toks = words(sample)
if not toks:
return "unk"
tr = sum(t in TR_WORDS for t in toks) + 3 * sum(c in TR_CHARS for c in sample) / max(1, len(sample) / 100)
en = sum(t in EN_WORDS for t in toks)
if tr > en and tr > 2:
return "tr"
if en >= 2 or (len(toks) < 30 and sample.isascii()):
return "en"
return "other"
def gopher_quality(text: str):
"""A subset of the Gopher (Rae et al., 2021) document quality rules. Returns list of failures."""
fails = []
toks = text.split()
n = len(toks)
if n < 50:
fails.append("too_short")
if n > 100_000:
fails.append("too_long")
if n:
mean = sum(len(t) for t in toks) / n
if not 3 <= mean <= 10:
fails.append("mean_word_length")
if sum(t.startswith(("#", "...")) or t.endswith("…") for t in toks) / n > 0.1:
fails.append("symbol_ratio")
alpha = sum(any(c.isalpha() for c in t) for t in toks) / n
if alpha < 0.8:
fails.append("non_alpha_words")
lines = [l for l in text.splitlines() if l.strip()]
if lines:
if sum(l.lstrip().startswith(("•", "-", "*")) for l in lines) / len(lines) > 0.9:
fails.append("bullet_lines")
if sum(l.rstrip().endswith(("...", "…")) for l in lines) / len(lines) > 0.3:
fails.append("ellipsis_lines")
dup = 1 - len(set(lines)) / len(lines)
if dup > 0.3:
fails.append("duplicate_lines")
return fails
def approx_tokens(text: str) -> int:
"""~4 characters per token for English/code; used only where an exact count is too costly."""
return max(1, round(len(text) / 4))