Spaces:
Running
Running
File size: 20,710 Bytes
16dc556 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 | """Field archetypes: clean-value generators + matched corruptors.
Each archetype produces clean values in the SAME canonical/typed representation
that scrubdata.executor outputs, and a `corrupt()` that dirties a clean column
while returning the exact ground-truth column-operations. Designed so that
executor(dirty, ops) == clean (verified downstream).
"""
from __future__ import annotations
import random
from . import vocab as V
# ---- shared corruption helpers ----------------------------------------------
DISGUISED = ["N/A", "na", "-", "--", "null", "None", "?", "#N/A", "TBD",
"empty", "(empty)", "n/a", "NULL", "none", "unknown"]
def _add_whitespace(rng: random.Random, s: str) -> str:
choice = rng.random()
if choice < 0.4:
return " " * rng.randint(1, 3) + s
if choice < 0.7:
return s + " " * rng.randint(1, 3)
# doubled internal space
parts = s.split(" ")
if len(parts) > 1:
i = rng.randrange(len(parts) - 1)
parts[i] = parts[i] + " "
return " ".join(parts)
return " " + s + " "
def _inject_disguised_nulls(rng: random.Random, values, clean, p=0.12):
"""Randomly turn some cells into disguised-null tokens; clean value = None."""
used = False
out_dirty, out_clean = [], []
for d, c in zip(values, clean):
if rng.random() < p:
out_dirty.append(rng.choice(DISGUISED))
out_clean.append(None)
used = True
else:
out_dirty.append(d)
out_clean.append(c)
return out_dirty, out_clean, used
# ---- archetypes --------------------------------------------------------------
class Field:
semantic_type = "text"
names: list[str] = []
def gen_clean(self, rng: random.Random, n: int):
raise NotImplementedError
def corrupt(self, rng: random.Random, clean):
"""Return (dirty_values, clean_values, ops, issues)."""
raise NotImplementedError
class NameField(Field):
semantic_type = "text"
names = ["name", "full_name", "customer", "contact", "rep"]
FIRST = ["Alice", "Bob", "Carol", "David", "Eve", "Frank", "Grace", "Heidi",
"Ivan", "Judy", "Karl", "Lena", "Mona", "Omar", "Priya", "Sara"]
LAST = ["Johnson", "Smith", "Diaz", "Lee", "Adams", "Moore", "Park", "Cruz",
"Petrov", "Wong", "Brandt", "Fischer", "Ali", "Khan", "Novak", "Reyes",
"O'Brien", "D'Angelo", "Saint-Clair", "Smith-Jones", "N'Diaye"]
def gen_clean(self, rng, n):
return [f"{rng.choice(self.FIRST)} {rng.choice(self.LAST)}" for _ in range(n)]
def corrupt(self, rng, clean):
dirty = [_add_whitespace(rng, c) if rng.random() < 0.5 else c for c in clean]
ops = [{"op": "strip_whitespace",
"rationale": "Trimmed leading/trailing and doubled spaces."}]
issues = ["whitespace"]
# high-cardinality regime: unicode punctuation artifacts (curly quotes, long
# dashes, NBSP). Inverse of executor._PUNCT_MAP -> execution-verified.
if rng.random() < 0.45:
punct = False
for i, v in enumerate(dirty):
if rng.random() < 0.35:
w = v.replace("'", "β").replace("-", "β")
if " " in w and rng.random() < 0.3:
k = w.rindex(" ")
w = w[:k] + "Β " + w[k + 1:]
if w != v:
dirty[i] = w
punct = True
if punct:
ops.append({"op": "normalize_punctuation",
"rationale": "Normalized curly quotes / long dashes / "
"NBSP artifacts to plain ASCII."})
issues.append("unicode_punctuation")
return dirty, clean, ops, issues
class CompanyField(NameField):
names = ["company", "organization", "account", "employer"]
POOL = ["Acme Inc", "Globex", "Initech", "Umbrella", "Soylent Corp", "Hooli",
"Vehement", "Stark Industries", "Wonka Co", "Cyberdyne",
"O'Reilly & Sons", "Day-Lewis Group", "L'Atelier Co"]
def gen_clean(self, rng, n):
return [rng.choice(self.POOL) for _ in range(n)]
class EmailField(Field):
semantic_type = "email"
names = ["email", "email_address", "contact_email"]
def gen_clean(self, rng, n):
out = []
for _ in range(n):
user = "".join(rng.choice("abcdefghijklmnop") for _ in range(rng.randint(4, 7)))
dom = rng.choice(["example.com", "mail.com", "corp.io", "test.org"])
out.append(f"{user}@{dom}")
return out
def corrupt(self, rng, clean):
dirty = []
for c in clean:
v = c.upper() if rng.random() < 0.5 else c
if rng.random() < 0.4:
v = _add_whitespace(rng, v)
dirty.append(v)
ops = [{"op": "normalize_email",
"rationale": "Lowercased and trimmed email addresses."}]
return dirty, clean, ops, ["casing", "whitespace"]
class VocabField(Field):
"""Categorical column backed by a real vocabulary (canonical -> aliases).
LOW-card mode (default): draws a FEW canonicals (every surface shows in the
sample). HIGH-card mode (high_card=True): draws MANY (min_card..max_card, e.g.
30..80) real canonicals with a DOMINANT-canonical long-tailed row distribution
and single-char-substitution typos in the tail β replicating the hospital
birmingham(75) + birminghxm(1) regime. Both corrupt() and record surface->
canonical so canonicalize_categories recovers the clean value (self-verified)."""
def __init__(self, names, semantic_type, entries, max_card=5, min_card=2,
high_card=False, typo_p=0.13):
self.names = names
self.semantic_type = semantic_type
self.entries = entries
self._canonicals = list(entries)
self.max_card = max_card
self.min_card = min_card
self.high_card = high_card
self.typo_p = typo_p
def _choose(self, rng):
lo = max(2, min(self.min_card, len(self._canonicals)))
hi = min(self.max_card, len(self._canonicals))
k = rng.randint(min(lo, hi), hi)
return rng.sample(self._canonicals, k)
def _gen_rows(self, rng, n):
"""Long-tailed row draw: a few dominant canonicals carry most of the mass,
the rest form a sparse tail (where typo surfaces land as rare singletons).
Falls back to uniform for low-card columns."""
chosen = self._chosen
if not self.high_card or len(chosen) < 6:
return [rng.choice(chosen) for _ in range(n)]
# Zipf-like weights: a couple of dominant values, steeply decaying tail.
order = list(chosen)
rng.shuffle(order)
weights = [1.0 / ((i + 1) ** 1.6) for i in range(len(order))]
# Boost the single top canonical so a clear dominant emerges (birmingham 75).
weights[0] *= 3.0
return rng.choices(order, weights=weights, k=n)
def gen_clean(self, rng, n):
self._chosen = self._choose(rng)
return self._gen_rows(rng, n)
def _surface_for(self, rng, c, force_typo):
"""One dirty surface for canonical c. force_typo guarantees a single-char
substitution typo (rare-tail birminghxm regime)."""
aliases = self.entries.get(c, [])
if force_typo:
s = V.make_substitution_typo(rng, c)
return s
return V.make_surface(rng, c, aliases, typo_p=self.typo_p)
def corrupt(self, rng, clean):
# Decide which canonicals get a guaranteed single-char typo surface (high-card
# only): a controlled fraction of the present canonicals, applied to ONE of
# their occurrences so it lands as a rare tail singleton.
present = list(dict.fromkeys(clean))
forced_typo_canon = set()
if self.high_card:
frac = rng.uniform(0.3, 0.6)
k = max(1, int(len(present) * frac))
forced_typo_canon = set(rng.sample(present, min(k, len(present))))
# Reserve, per forced canonical, exactly one row index to carry the typo.
forced_slot = {}
if forced_typo_canon:
for canon in forced_typo_canon:
idxs = [i for i, c in enumerate(clean) if c == canon]
if idxs:
forced_slot[rng.choice(idxs)] = canon
# Build mapping collision-safely: a surface may only map to ONE canonical, and
# a surface that equals some canonical's clean form must not be remapped.
# Reserve all clean canonical strings as "do not remap" keys.
reserved = {str(c).strip() for c in present}
mapping = {}
dirty, ws = [], False
for i, c in enumerate(clean):
force = i in forced_slot
for _attempt in range(4):
s = self._surface_for(rng, c, force_typo=force)
key = str(s).strip()
if key == str(c).strip():
break # already canonical surface, no mapping needed
# Skip surfaces that collide with another canonical, or that some
# other canonical already claims (would make the mapping ambiguous).
if key in reserved:
s = c # ambiguous -> fall back to clean (still verifies)
break
if key in mapping and mapping[key] != c:
s = c # collision with a different canonical's surface
break
break
key = str(s).strip()
if key != str(c).strip() and key not in reserved:
mapping[key] = c
cell = s
# whitespace noise (less often on high-card to keep the tail clean)
if rng.random() < (0.12 if self.high_card else 0.25):
cell = _add_whitespace(rng, s)
ws = True
dirty.append(cell)
ops, issues = [], ["inconsistent_categories", "casing"]
if ws: # strip first so canonicalize sees the bare surface (executor order)
ops.append({"op": "strip_whitespace",
"rationale": "Trimmed surrounding/doubled spaces."})
issues.append("whitespace")
if mapping:
ops.append({"op": "canonicalize_categories", "mapping": mapping,
"rationale": f"Unified {len(mapping)} variant spelling(s) "
f"into canonical labels."})
return dirty, clean, ops, issues
class StatusField(VocabField):
"""Like VocabField but picks a fresh status/category value-set each example."""
def __init__(self):
super().__init__(
names=["status", "stage", "tier", "segment", "state", "payment_status"],
semantic_type="categorical", entries={}, max_card=4)
def gen_clean(self, rng, n):
self.entries = rng.choice(V._STATUS_SETS)
self._canonicals = list(self.entries)
self._chosen = self._choose(rng)
return self._gen_rows(rng, n)
class CurrencyField(Field):
semantic_type = "currency"
names = ["amount", "revenue", "price", "deal_size", "cost"]
def gen_clean(self, rng, n):
return [round(rng.uniform(50, 9000), 2) for _ in range(n)]
def _fmt(self, rng, x: float) -> str:
neg = x < 0
a = abs(x)
style = rng.random()
if style < 0.4:
s = f"${a:,.2f}"
elif style < 0.7 and a == int(a):
s = f"{int(a):,d}" # grouped integer β only when no cents to lose
else: # EU style (comma decimal) β always preserves 2 decimals
s = f"{a:,.2f}".replace(",", "X").replace(".", ",").replace("X", ".")
return f"({s})" if neg else s
def corrupt(self, rng, clean):
dirty = [self._fmt(rng, c) for c in clean]
dirty, clean2, used_null = _inject_disguised_nulls(rng, dirty, clean)
ops, issues = [], ["numeric_stored_as_text", "currency_symbols"]
if used_null:
ops.append({"op": "normalize_disguised_nulls",
"rationale": "Converted N/A, '-', 'null' etc. to true missing."})
issues.append("disguised_nulls")
ops.append({"op": "parse_currency",
"rationale": "Stripped currency symbols/grouping; parsed to number."})
return dirty, clean2, ops, issues
class DateField(Field):
semantic_type = "date"
names = ["signup_date", "created_at", "close_date", "date", "order_date"]
def gen_clean(self, rng, n):
out = []
for _ in range(n):
y, m, d = 2023, rng.randint(1, 12), rng.randint(1, 28)
out.append(f"{y:04d}-{m:02d}-{d:02d}")
return out
def _fmt(self, rng, iso: str) -> str:
y, m, d = iso.split("-")
months = ["Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug",
"Sep", "Oct", "Nov", "Dec"]
style = rng.random()
if style < 0.3:
return iso
if style < 0.55:
return f"{int(m)}/{int(d)}/{y}" # US slash (m<=12, d<=28 -> unambiguous-ish)
if style < 0.8:
return f"{int(d)} {months[int(m)-1]} {y}" # 5 Jan 2023
# Excel serial
import datetime
base = datetime.date(1899, 12, 30)
serial = (datetime.date(int(y), int(m), int(d)) - base).days
return str(serial)
def corrupt(self, rng, clean):
dirty = [self._fmt(rng, c) for c in clean]
ops = [{"op": "parse_date",
"rationale": "Unified mixed date formats to ISO YYYY-MM-DD."}]
return dirty, clean, ops, ["mixed_date_formats"]
class BooleanField(Field):
semantic_type = "boolean"
names = ["is_active", "subscribed", "verified", "opted_in"]
TRUE = ["Yes", "Y", "TRUE", "true", "1", "T"]
FALSE = ["No", "N", "FALSE", "false", "0", "F"]
def gen_clean(self, rng, n):
return [rng.random() < 0.5 for _ in range(n)]
def corrupt(self, rng, clean):
dirty = [rng.choice(self.TRUE if c else self.FALSE) for c in clean]
ops = [{"op": "standardize_boolean",
"rationale": "Mapped Yes/Y/1/TRUE β true, No/N/0/FALSE β false."}]
return dirty, clean, ops, ["inconsistent_booleans"]
class PhoneField(Field):
semantic_type = "phone"
names = ["phone", "phone_number", "mobile", "contact_number"]
def gen_clean(self, rng, n):
# Canonical = executor's output for a plain 10-digit US number.
out, self._digits = [], []
for _ in range(n):
d = "".join(str(rng.randint(0, 9)) for _ in range(10))
d = "5" + d[1:] # keep it phone-ish
self._digits.append(d)
out.append(f"({d[0:3]}) {d[3:6]}-{d[6:]}")
return out
def corrupt(self, rng, clean):
dirty = []
for d in self._digits:
style = rng.random()
if style < 0.25:
dirty.append(f"{d[0:3]}.{d[3:6]}.{d[6:]}")
elif style < 0.5:
dirty.append(f"{d[0:3]}-{d[3:6]}-{d[6:]}")
elif style < 0.75:
dirty.append(d)
else:
dirty.append(f"({d[0:3]}){d[3:6]}-{d[6:]}")
ops = [{"op": "standardize_phone",
"rationale": "Standardized phone formatting."}]
return dirty, clean, ops, ["inconsistent_formats"]
class PercentField(Field):
semantic_type = "percent"
names = ["rate", "discount", "completion", "margin", "growth", "conversion"]
def gen_clean(self, rng, n):
self._pct = [round(rng.uniform(0, 100), 1) for _ in range(n)]
return [p / 100 for p in self._pct]
def corrupt(self, rng, clean):
dirty = [f"{p}%" for p in self._pct]
ops = [{"op": "parse_percent", "rationale": "Parsed percent text to a fraction."}]
return dirty, clean, ops, ["numeric_stored_as_text"]
ARCHETYPES: list[Field] = [
NameField(), CompanyField(), EmailField(), PercentField(),
VocabField(["country", "nation", "country_name"], "country", V.country_vocab(), max_card=5),
VocabField(["state", "province", "region"], "state", V.state_vocab(), max_card=5),
VocabField(["currency", "currency_code", "ccy"], "categorical", V.currency_vocab(), max_card=4),
VocabField(["city", "location", "hq_city"], "city", V.city_vocab(), max_card=5),
VocabField(["department", "dept", "team"], "categorical", V.department_vocab(), max_card=4),
VocabField(["job_title", "title", "role", "position"], "categorical", V.job_title_vocab(), max_card=4),
# real O*NET occupations (alternate title -> canonical, CC BY 4.0): 1,016 canonicals
*([VocabField(["job_title", "occupation", "role"], "categorical",
V._cached("onet", lambda: V._alias_file("onet_jobtitle_aliases.jsonl", limit=1016)),
max_card=5),
VocabField(["job_title", "occupation"], "categorical",
V._cached("onet", lambda: V._alias_file("onet_jobtitle_aliases.jsonl", limit=1016)),
min_card=25, max_card=60, high_card=True)]
if V._alias_file("onet_jobtitle_aliases.jsonl", limit=2) else []),
# real nickname->formal first names (Bill -> William; Apache-2.0)
*([VocabField(["first_name", "given_name", "contact_first"], "categorical",
V.nickname_vocab(), max_card=5),
VocabField(["first_name", "given_name"], "categorical",
V.nickname_vocab(), min_card=25, max_card=60, high_card=True)]
if V.nickname_vocab() else []),
# ToughTables gold-anchored entity misspellings (SemTab 2T, CC-BY-4.0): 49.6k real
# variant aliases across people/films/places β the grouped-entity regime
*([VocabField(["name", "entity", "person", "artist"], "categorical",
V._cached("tt", lambda: V._alias_file("toughtables_aliases.jsonl", limit=3000)),
max_card=5),
VocabField(["name", "entity", "player"], "categorical",
V._cached("tt", lambda: V._alias_file("toughtables_aliases.jsonl", limit=3000)),
min_card=25, max_card=60, high_card=True)]
if V._alias_file("toughtables_aliases.jsonl", limit=2) else []),
# RxNorm prescribable drugs (public domain): synonym/TTY variants -> ingredient
*([VocabField(["drug", "medication", "drug_name", "prescription"], "categorical",
V._cached("rxnorm", lambda: V._alias_file("rxnorm_aliases.jsonl", limit=1500)),
max_card=5),
VocabField(["drug", "medication"], "categorical",
V._cached("rxnorm", lambda: V._alias_file("rxnorm_aliases.jsonl", limit=1500)),
min_card=25, max_card=60, high_card=True)]
if V._alias_file("rxnorm_aliases.jsonl", limit=2) else []),
# MusicBrainz search-hint aliases (CC0): community-recorded artist misspellings
*([VocabField(["artist", "performer", "band", "composer"], "categorical",
V._cached("mbhint", lambda: V._alias_file("musicbrainz_hint_aliases.jsonl", limit=2000)),
max_card=5),
VocabField(["artist", "performer"], "categorical",
V._cached("mbhint", lambda: V._alias_file("musicbrainz_hint_aliases.jsonl", limit=2000)),
min_card=25, max_card=60, high_card=True)]
if V._alias_file("musicbrainz_hint_aliases.jsonl", limit=2) else []),
VocabField(["industry", "sector", "vertical"], "categorical", V.industry_vocab(), max_card=4),
# real Wikidata companies (alias -> canonical: 'AB InBev' -> 'Anheuser-Busch InBev')
*([VocabField(["company", "vendor", "account", "supplier"], "categorical",
V.company_vocab(), max_card=5),
VocabField(["company", "vendor", "account"], "categorical",
V.company_vocab(), min_card=25, max_card=60, high_card=True)]
if V.company_vocab() else []),
# real ROR organizations (alias/acronym -> canonical): both low-card and the
# hospital-style high-cardinality long-tail regime. Skipped if harvest absent.
*([VocabField(["organization", "institution", "affiliation", "employer"], "categorical",
V.org_vocab(), max_card=5),
VocabField(["organization", "institution", "affiliation"], "categorical",
V.org_vocab(), min_card=25, max_card=60, high_card=True)]
if V.org_vocab() else []),
VocabField(["unit", "uom", "measure_unit"], "categorical", V.unit_vocab(), max_card=4),
StatusField(),
CurrencyField(), DateField(), BooleanField(), PhoneField(),
]
|