Datasets:
File size: 21,151 Bytes
e88fd67 | 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 | #!/usr/bin/env python
"""Corpus-wide regeneration of `caption_general`.
TWO defects are repaired in ONE pass over the index parquets.
Defect 1 -- the emotion clause was gated on an ABSOLUTE raw threshold (emo_thr=1.0).
The 40 Empathic-Insight heads are not on a common scale: emo_Interest has median
2.082 and 0.0 % zeros, emo_Infatuation median -0.017 and 87.7 % zeros. The absolute
gate therefore named Interest on 94.8 % of clips and Sadness on almost none -- it was
reporting the scale of the head, not the emotion of the clip.
FIX: name an emotion when it is in the top 10 % FOR THAT EMOTION against the pooled
tie-aware mid-rank ECDF in emonorm/out/capnorm.npz (132,833,726 rows), max 3 named.
A clip that clears nothing says "no dominant emotion".
Defect 2 -- the GEND and BKGN ordinal ladders ran BACKWARDS.
High vn_GEND is MASCULINE (+0.842 with chest resonance, -0.484 with head resonance)
but was rendered "feminine". High vn_BKGN is CLEANER (+0.786 with recording quality,
+0.29 with the independent qual_background_quality head) but was rendered "noisy".
caption2.py was fixed 2026-08-22 (md5 ec55b223) but the corpus rows were never redone.
WHY THIS IS DRIVEN FROM THE BUCKET COLUMN, NOT BY FLIPPING THE STRING
STATE AS OF 2026-08-23, AFTER THIS PASS: every tree is corrected. All 48,556 shards /
165,516,420 rows carry the corrected polarity and a `caption_general_v1` column holding
the previous text. Do not read the paragraph below as a description of the corpus today.
AS MEASURED BEFORE THE PASS (2026-08-22, the reason for this design): part of the corpus
was ALREADY correct. The `laion-tts-annotated-v1-reann` tree (vprof_base +
vprof_repaired, 28.2 M rows) had been regenerated after the caption2.py fix and measured
100 % new-polarity, while all nine live-tree datasets measured 100 % old-polarity. A
blind string flip would therefore have re-inverted those 28.2 M rows. Instead the correct
tag is recomputed from vn_GEND_bucket / vn_BKGN_bucket and substituted, which is
IDEMPOTENT: running this twice is the same as running it once -- which is also why this
file is safe to re-run now that the whole corpus is already correct.
ONLY three clauses may change. The caption is split on "; ", the GEND token is replaced
inside clause 0, the BKGN token inside the recording clause, and the "reads as" clause is
replaced wholesale. Every other clause is carried across byte-identical and that is
asserted per shard, so delivery, timbre, speech, affect, style, recording quality, the
explicit flag, burst handling, genuineness, blend, duration and language cannot move.
"""
import glob, json, os, re, socket, sys, traceback
import numpy as np
import pyarrow as pa
import pyarrow.parquet as pq
NB = "/e/data1/datasets/playground/mmlaion/schuhmann1/dramabox"
CAPNORM = f"{NB}/emonorm/out/capnorm.npz"
NAME_RANK = None
U_FLOOR = 0.90
TOP_N = 3
NBIN = 4096
NONE_CLAUSE = "no dominant emotion"
V1COL = "caption_general_v1"
# corrected ladders, identical to caption2.py md5 ec55b223 (low bucket -> feminine / noisy)
GEND = ["strongly feminine", "feminine", "somewhat feminine", "androgynous",
"somewhat masculine", "masculine", "strongly masculine"]
BKGN = ["very noisy background", "noisy background", "some background noise",
"quiet background", "no background noise"]
# every token either ladder can have produced, longest first so "very noisy background"
# is matched before "noisy background".
_G_ALL = sorted(set(GEND), key=len, reverse=True)
_B_ALL = sorted(set(BKGN), key=len, reverse=True)
G_RE = re.compile(r"\b(" + "|".join(map(re.escape, _G_ALL)) + r")\b")
B_RE = re.compile(r"\b(" + "|".join(map(re.escape, _B_ALL)) + r")\b")
READS_RE = re.compile(r"^reads as\b", re.I)
EXPL_TAGS = {"clean content", "mildly explicit content", "explicit content"}
RCQL_TAGS = ["very poor recording", "poor recording", "below-average recording",
"average recording", "good recording", "very good recording",
"studio-grade recording"]
RCQL_RE = re.compile(r"\b(" + "|".join(map(re.escape,
sorted(RCQL_TAGS, key=len, reverse=True))) + r")\b")
def pretty(e):
return e[4:].replace("_", " ").replace("/", " or ").lower()
# Alphabetical rank of each emotion's DISPLAY name, in capnorm field order. Used only to
# break genuine exact ties, deterministically and independently of input order.
def _name_rank(fields):
names = [pretty(f) for f in fields]
order = sorted(range(len(names)), key=lambda i: names[i])
rank = np.empty(len(names), np.int64)
for r, i in enumerate(order):
rank[i] = r
return rank
class Norm:
"""Pooled tie-aware mid-rank ECDF, loaded from capnorm.npz. Not re-fitted."""
def __init__(self, path=CAPNORM):
z = np.load(path, allow_pickle=True)
self.fields = [str(x) for x in z["fields"]]
h = z["hist"].astype(np.float64)
n = np.maximum(h.sum(1, keepdims=True), 1.0)
below = np.cumsum(h, axis=1) - h
# float64, NOT float32. Casting the mid-rank table to float32 collapses
# genuinely different percentiles onto one value: emo_Relief 0.99780922730421640
# and emo_Contentment 0.99780920848369492 both become 0.99780923128128052, so the
# renderer saw a tie that does not exist and resolved it arbitrarily. Measured
# cost of that artefact: 10 disagreements with the templates in 361,277 rows, two
# of which changed WHICH emotion took the third slot. float64 also matches
# capgate.py and caption_render.py, which never cast.
self.tab = (below + 0.5 * h) / n
self.lo = z["lo"].astype(np.float64)
self.hi = z["hi"].astype(np.float64)
self.n = int(z["n"][0])
self.w = (self.hi - self.lo) / NBIN
global NAME_RANK
NAME_RANK = _name_rank(self.fields)
def u(self, X):
k = np.floor((np.asarray(X, np.float64) - self.lo) / self.w) + 1.0
np.clip(k, 0, NBIN + 1, out=k)
k = k.astype(np.int32)
Y = np.empty(k.shape, np.float64)
for d in range(k.shape[1]):
Y[:, d] = self.tab[d][k[:, d]]
return Y
def split_clauses(cap):
"""'a; b; c.' -> (['a','b','c'], True) trailing-dot flag preserved."""
s = cap.rstrip()
dot = s.endswith(".")
if dot:
s = s[:-1]
return [c.strip() for c in s.split(";")], dot
def join_clauses(parts, dot):
return "; ".join(parts) + ("." if dot else "")
def fix_one(cap, gb, bb, names, stat):
"""Return the regenerated caption. Only clause 0 (GEND), the recording clause
(BKGN) and the emotion clause may differ; everything else is carried across."""
if not cap:
return cap
parts, dot = split_clauses(cap)
if not parts:
return cap
# ---- GEND, inside clause 0 only ----
if gb is not None and gb == gb: # not NaN
want = GEND[min(max(int(gb), 0), 6)]
m = G_RE.search(parts[0])
if m:
if m.group(1) != want:
stat["gend_changed"] += 1
parts[0] = parts[0][:m.start()] + want + parts[0][m.end():]
# the article depends on the first letter of the who-phrase; recompute it.
# (no ladder term starts with a vowel except "androgynous", which is the
# self-symmetric middle bucket, so this is a no-op in practice -- asserted.)
mm = re.match(r"^(An?) (.+)$", parts[0])
if mm:
art = "An" if mm.group(2)[:1].lower() in "aeiou" else "A"
if art != mm.group(1):
stat["article_changed"] += 1
parts[0] = f"{art} {mm.group(2)}"
# ---- BKGN, inside whichever clause carries a background token ----
if bb is not None and bb == bb:
want = BKGN[min(max(int(bb), 0), 4)]
for i, p in enumerate(parts):
m = B_RE.search(p)
if m:
if m.group(1) != want:
stat["bkgn_changed"] += 1
parts[i] = p[:m.start()] + want + p[m.end():]
break
# ---- emotion clause ----
# Always remove any existing emotion clause and re-insert at the CANONICAL
# position, so the result is independent of whether the old caption happened to
# carry one. caption2.py emits, in order:
# who; delivery; timbre; speech; affect; EMOTION; style; recording; explicit;
# bursts; genuineness; blend; tail
# so the emotion clause belongs immediately before the first of
# {style, recording, explicit, bursts, genuineness}. An earlier version inserted
# it just before "genuineness", which put it AFTER the recording and burst
# clauses on the 12.5 M rows that previously had no emotion clause at all -- same
# content, wrong slot, and enough to make caption_general disagree with a fresh
# caption_clausal render. Removing first also makes this idempotent.
new = ("reads as " + ", ".join(names)) if names else NONE_CLAUSE
had = [i for i, p in enumerate(parts) if READS_RE.match(p) or p == NONE_CLAUSE]
prev = parts[had[0]] if had else None
for i in reversed(had):
parts.pop(i)
def _is_anchor(c):
cl = c.lower()
if cl.startswith(("style: ", "genuineness", "contains vocal bursts")):
return True
if c in EXPL_TAGS:
return True
return bool(B_RE.search(c)) or bool(RCQL_RE.search(c))
pos = next((i for i, c in enumerate(parts) if _is_anchor(c)), len(parts))
parts.insert(pos, new)
if prev is None:
stat["emo_inserted"] += 1
elif prev != new:
stat["emo_changed"] += 1
return join_clauses(parts, dot)
def already_done(path):
"""A shard carrying the v1 column has been through this pass. Cheap metadata read."""
try:
return V1COL in pq.ParquetFile(path).schema_arrow.names
except Exception:
return False
def process(path, norm, floor=U_FLOOR, top=TOP_N, dry=False, outdir=None, attempts=4):
"""Rewrite one shard, retrying transient filesystem faults. Never raises.
RETRY EXISTS BECAUSE OF A MEASURED FAULT, not as decoration. On the first
corpus-wide run (16 nodes x 36 workers = 576 processes creating and renaming
files inside the same directories) ~11 % of shards failed with the freshly
written temp file reported missing -- ENOENT from pq.ParquetFile(tmp) or from
os.replace(tmp, path). The identical workload on ONE node (600 shards, 36
workers) failed 0 times, so this is cross-node metadata contention on the
parallel filesystem, not a defect in the shard.
Two things make it safe to simply retry: the temp name is now unique per
process, so nothing can collide; and every failure happens strictly BEFORE
os.replace, so the original shard is still the original shard. A shard that
exhausts its attempts is left untouched and reported, never half-written.
"""
last = None
for k in range(attempts):
st = _process_once(path, norm, floor, top, dry, outdir)
if st.get("ok") or dry:
if k:
st["retries"] = k
return st
last = st
e = st.get("err", "")
transient = ("No such file or directory" in e or "Failed to open local file" in e
or "magic bytes" in e or "File too short" in e
or "smaller than the minimum file footer" in e
or "Couldn't deserialize thrift" in e)
if not transient:
return st
time.sleep(0.4 * (k + 1) + random.random() * 0.4)
last["retries"] = attempts
return last
def _process_once(path, norm, floor=U_FLOOR, top=TOP_N, dry=False, outdir=None):
"""Rewrite one shard. Returns a stats dict. Never raises."""
st = dict(path=path, ok=0, rows=0, rows_after=0, uid_n=0, uid_uniq=0,
gend_changed=0, bkgn_changed=0, emo_changed=0, emo_inserted=0,
article_changed=0, none=0, named=0, err="", had_v1=0,
emo_hist={}, emo_count={}, before_count={}, before_hist={},
before_none=0, bytes_before=0, bytes_after=0, other_clause_moved=0)
tmp = None
try:
# ONE authoritative read. Previously the schema came from a separate
# pq.ParquetFile(path) open and the data from pq.read_table(path); a
# concurrent job replacing the file between those two reads made
# `names_in` describe the OLD file while `tab` was the NEW one, so the
# code appended caption_general_v1 to a table that already had it. That
# produced 22 shards with a duplicated column. Reading the schema off the
# very table being transformed makes the race impossible.
# ParquetFile.read(), NOT pq.read_table(): read_table goes through the
# pyarrow.dataset layer, which resolves columns by FieldRef.Name and dies
# with "Multiple matches for FieldRef.Name(caption_general_v1)" on exactly
# the duplicated-column shards this function is meant to self-heal. The
# file-level reader has no such name resolution and reads duplicates fine.
tab = pq.ParquetFile(path).read()
names_in = list(tab.schema.names)
st["rows"] = tab.num_rows
n = tab.num_rows
if "caption_general" not in names_in:
st["err"] = "no caption_general"
return st
# self-heal any shard already damaged by that race: keep the FIRST
# caption_general_v1 (the true original caption) and drop the later one
# (which holds an already-rewritten caption and is not a baseline).
while names_in.count(V1COL) > 1:
last = len(names_in) - 1 - names_in[::-1].index(V1COL)
tab = tab.remove_column(last)
names_in = list(tab.schema.names)
st["dup_v1_dropped"] = st.get("dup_v1_dropped", 0) + 1
# ---- pre-checks ----
if "uid" in names_in:
uid = tab.column("uid").to_pylist()
st["uid_n"] = len(uid)
st["uid_uniq"] = len(set(uid))
cap = tab.column("caption_general").to_pylist()
st["had_v1"] = int(V1COL in names_in)
gb = tab.column("vn_GEND_bucket").to_pylist() if "vn_GEND_bucket" in names_in else [None] * n
bb = tab.column("vn_BKGN_bucket").to_pylist() if "vn_BKGN_bucket" in names_in else [None] * n
# ---- emotion percentiles ----
X = np.empty((n, len(norm.fields)), np.float64)
for j, f in enumerate(norm.fields):
if f in names_in:
col = tab.column(f).to_numpy(zero_copy_only=False).astype(np.float64)
else:
col = np.zeros(n)
X[:, j] = np.nan_to_num(col, nan=0.0, posinf=0.0, neginf=0.0)
U = norm.u(X)
# EXPLICIT tie-break: descending percentile, then ASCENDING EMOTION NAME.
# Not merely a stable sort -- a stable sort preserves *input* order, which a
# consumer of the corpus cannot see or reason about. Ordering genuine ties by
# display name is a rule anyone can reproduce from the caption alone, and it is
# the same rule caption_render.py and capgate.py now use, so the column and the
# 16 templates cannot disagree. np.lexsort takes the LAST key as primary.
order = np.lexsort((NAME_RANK[None, :].repeat(U.shape[0], 0), -U), axis=1)[:, :top]
# ---- what the OLD caption named, for the before/after report ----
# If v1 already exists this shard was rewritten by an earlier (cancelled) run, so
# the TRUE baseline is v1, not the current caption_general. Reading the wrong one
# would make the before/after report describe the fix against itself.
base = (tab.column(V1COL).to_pylist() if V1COL in names_in else cap)
bcount, bhist = {}, {}
for c in base:
if not c:
continue
names_old = []
for p in split_clauses(c)[0]:
if READS_RE.match(p):
names_old = [x.strip() for x in p[len("reads as"):].split(",") if x.strip()]
break
bhist[len(names_old)] = bhist.get(len(names_old), 0) + 1
if not names_old:
st["before_none"] += 1
for x in names_old:
bcount[x] = bcount.get(x, 0) + 1
st["before_count"] = bcount
st["before_hist"] = bhist
newcap = []
ecount = {}
ehist = {}
for i in range(n):
sel = [norm.fields[j] for j in order[i] if U[i, j] >= floor]
nm = [pretty(e) for e in sel]
ehist[len(nm)] = ehist.get(len(nm), 0) + 1
if nm:
st["named"] += 1
for x in nm:
ecount[x] = ecount.get(x, 0) + 1
else:
st["none"] += 1
newcap.append(fix_one(cap[i], gb[i], bb[i], nm, st))
st["emo_count"] = ecount
st["emo_hist"] = ehist
# ---- requirement 4: assert nothing else moved ----
moved = 0
for o, nw in zip(cap, newcap):
if not o or not nw:
continue
po, _ = split_clauses(o)
pn, _ = split_clauses(nw)
# drop the emotion / none clause from both sides, then the remaining
# clause lists must differ only where GEND/BKGN tokens live
fo = [c for c in po if not READS_RE.match(c) and c != NONE_CLAUSE]
fn = [c for c in pn if not READS_RE.match(c) and c != NONE_CLAUSE]
if len(fo) != len(fn):
moved += 1
continue
for a, b in zip(fo, fn):
if a == b:
continue
if G_RE.sub("", a) == G_RE.sub("", b) or B_RE.sub("", a) == B_RE.sub("", b):
continue
moved += 1
break
st["other_clause_moved"] = moved
if dry:
st["ok"] = 1
st["rows_after"] = n
st["_sample"] = [(cap[i], newcap[i]) for i in range(min(3, n))]
return st
# ---- write ----
if V1COL in names_in:
# already preserved on an earlier run: keep the ORIGINAL v1, do not overwrite
tab = tab.set_column(names_in.index("caption_general"), "caption_general",
pa.array(newcap, pa.string()))
else:
tab = tab.set_column(names_in.index("caption_general"), "caption_general",
pa.array(newcap, pa.string()))
tab = tab.append_column(V1COL, pa.array(cap, pa.string()))
st["bytes_before"] = os.path.getsize(path)
# UNIQUE tmp name. Two array jobs were once launched against the same shard list
# (1460437 by a sibling agent, 1460453 by me) and a shared "<path>.tmp-capfix"
# means two processes truncating the SAME file while each verifies it. The
# transform is deterministic so the output would agree, but a half-written file
# must never be a candidate for os.replace. Unique per process + a re-check that
# the source has not changed under us makes concurrent runs merely wasteful.
uniq = f"{os.getpid()}-{socket.gethostname()}-{os.urandom(4).hex()}"
tmp = ((outdir + "/" + os.path.basename(path)) if outdir
else f"{path}.tmp-capfix-{uniq}")
os.makedirs(os.path.dirname(tmp), exist_ok=True)
pq.write_table(tab, tmp, compression="zstd", compression_level=7,
use_dictionary=True, version="2.6")
# ---- verify BEFORE replacing ----
chk = pq.ParquetFile(tmp)
st["rows_after"] = chk.metadata.num_rows
if st["rows_after"] != st["rows"]:
os.remove(tmp)
st["err"] = f"row count {st['rows']} -> {st['rows_after']}"
return st
if "uid" in names_in:
u2 = pq.read_table(tmp, columns=["uid"]).column("uid").to_pylist()
if len(u2) != st["uid_n"] or len(set(u2)) != st["uid_uniq"]:
os.remove(tmp)
st["err"] = "uid check failed"
return st
if V1COL not in pq.ParquetFile(tmp).schema_arrow.names:
os.remove(tmp)
st["err"] = "v1 column missing after write"
return st
st["bytes_after"] = os.path.getsize(tmp)
if not outdir:
os.replace(tmp, path) # atomic; original stays intact until this instant
st["ok"] = 1
return st
except Exception as e:
st["err"] = f"{type(e).__name__}: {e}"
st["tb"] = traceback.format_exc()[-800:]
try:
if not outdir:
for junk in glob.glob(path + ".tmp-capfix.*"):
os.remove(junk)
except Exception:
pass
return st
finally:
# never leave a stale unique tmp behind if anything above bailed out
try:
if (not outdir) and tmp and os.path.exists(tmp):
os.remove(tmp)
except Exception:
pass
|