File size: 9,760 Bytes
80c9326 | 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 | """Tokenizer-training corpora from raw (unfiltered) FineWeb / FineWeb2.
Both tokenizer conditions are sampled from the SAME corpus family as the
model-training pools (data/fineweb.py's FineWeb-HQ / FineWeb2-HQ), just the
unfiltered releases -- HuggingFaceFW/fineweb (English) and
HuggingFaceFW/fineweb-2 (everyone else). This is a deliberate deviation from
ATLAS's literal MADLAD-400-trained tokenizer: it removes a tokenizer-corpus-
vs-model-corpus domain-mismatch confound that could otherwise hit AR/ZH
differently than DE/FR (MADLAD's non-Latin-script cleaning/LangID is less
consistent than FineWeb's), at the cost of no longer being a byte-for-byte
ATLAS replication for the starved tokenizer's source text.
starved -- ATLAS-style: T=100 temperature sampling (p_l ~ n_l^(1/100),
i.e. near-uniform) over ~419 languages -- English (raw FineWeb)
plus the ~418 largest FineWeb2 language-script configs by
volume. "Largest by volume" is the FineWeb2 analogue of how
MADLAD-400's own ~419-language "clean" set was itself
determined (languages with enough clean text to clear a
volume floor), so it preserves the same selection logic, just
applied to a different corpus.
destarved -- our 5 study languages only; per-language byte budgets scaled
by the FLORES+ byte premium so *content* (not bytes) is
uniform across languages.
Both stream parquet with column pruning (reusing data.fineweb's `_iter_texts`),
so only the `text` column is ever pulled.
"""
import json
import random
import urllib.request
from pathlib import Path
from ..langs import LANGS
from ..paths import MANIFEST_CACHE, TOK_CORPORA, ensure
from .fineweb import _iter_texts
FINEWEB_EN_REPO = "HuggingFaceFW/fineweb"
FINEWEB2_REPO = "HuggingFaceFW/fineweb-2"
N_STARVED_LANGS = 419 # matches MADLAD-400/ATLAS's ~419-language scale
# SentencePiece skips sentences longer than max_sentence_length (default 4192
# bytes); we pre-split long documents so no text is silently dropped.
MAX_LINE_BYTES = 4000
def _get_json(url: str) -> dict:
with urllib.request.urlopen(url, timeout=120) as r:
return json.loads(r.read().decode("utf-8"))
def _fineweb2_size_manifest(refresh: bool = False) -> dict[str, int]:
"""{config_name: num_bytes_original_files} for every FineWeb2 language-script config."""
cache = ensure(MANIFEST_CACHE) / "fineweb2_sizes.json"
if cache.exists() and not refresh:
return json.loads(cache.read_text())
d = _get_json(f"https://datasets-server.huggingface.co/size?dataset={FINEWEB2_REPO}")
sizes = {c["config"]: c["num_bytes_original_files"] for c in d["size"]["configs"]}
cache.write_text(json.dumps(sizes))
print(f"[tokcorpus] fetched FineWeb2 sizes for {len(sizes)} configs")
return sizes
def _fineweb_en_size(refresh: bool = False) -> int:
cache = ensure(MANIFEST_CACHE) / "fineweb_en_size.json"
if cache.exists() and not refresh:
return json.loads(cache.read_text())["bytes"]
d = _get_json(f"https://datasets-server.huggingface.co/size?dataset={FINEWEB_EN_REPO}&config=default")
n = d["size"]["config"]["num_bytes_original_files"]
cache.write_text(json.dumps({"bytes": n}))
return n
def select_starved_languages(n_langs: int = N_STARVED_LANGS) -> dict[str, int]:
"""{code: available_bytes} for the starved condition's language universe.
"en" (raw FineWeb) plus the (n_langs-1) largest FineWeb2 configs by volume.
Our other 4 study languages (de/fr/ar/zh) are always near the top of that
ranking on volume alone (verified: ranks 2-13 of 1314), so no forced
inclusion is needed.
"""
sizes = _fineweb2_size_manifest()
top = sorted(sizes.items(), key=lambda kv: -kv[1])[: n_langs - 1]
universe = {"en": _fineweb_en_size(), **dict(top)}
return universe
def _source_for_code(code: str) -> tuple[str, str]:
"""(repo, subdir) of raw FineWeb-family text for a language code.
`code` is either "en", one of our other 4 study codes, or a raw FineWeb2
language-script config name (a starved-only competing language).
"""
if code == "en":
return FINEWEB_EN_REPO, "data"
if code in LANGS:
code = LANGS[code].fineweb_subdir # study code -> FineWeb2 config name
return FINEWEB2_REPO, f"data/{code}/train"
def _list_parquet_files(repo: str, subdir: str) -> list[str]:
cache = ensure(MANIFEST_CACHE / "parquet_files") / f"{repo.replace('/', '__')}__{subdir.replace('/', '_')}.json"
if cache.exists():
return json.loads(cache.read_text())
from huggingface_hub import HfApi
api = HfApi()
files = [e.path for e in api.list_repo_tree(repo, subdir, repo_type="dataset", recursive=True)
if e.__class__.__name__ == "RepoFile" and e.path.endswith(".parquet")]
# dump-stratified sources (raw FineWeb's CC-MAIN-*/) interleave across dumps
# for a temporally representative sample; FineWeb2's per-language train/
# files have no such structure, so plain sort is enough.
by_dump: dict[str, list[str]] = {}
for f in sorted(files):
parts = f[len(subdir):].strip("/").split("/")
key = parts[0] if len(parts) > 1 and "CC-MAIN" in parts[0] else ""
by_dump.setdefault(key, []).append(f)
if len(by_dump) > 1:
out = []
for i in range(max(len(v) for v in by_dump.values())):
for d in sorted(by_dump):
if i < len(by_dump[d]):
out.append(by_dump[d][i])
files = out
else:
files = sorted(files)
cache.write_text(json.dumps(files))
return files
def _split_line(text: str):
"""Yield non-empty lines, splitting any line over MAX_LINE_BYTES at whitespace."""
for ln in text.split("\n"):
ln = ln.strip()
if not ln:
continue
enc = ln.encode("utf-8")
while len(enc) > MAX_LINE_BYTES:
cut = enc[:MAX_LINE_BYTES].rfind(b" ")
if cut < MAX_LINE_BYTES // 2:
cut = MAX_LINE_BYTES
while cut > 0 and (enc[cut] & 0xC0) == 0x80: # don't split mid UTF-8 char
cut -= 1
yield enc[:cut].decode("utf-8", errors="replace").strip()
enc = enc[cut:].lstrip()
if enc:
yield enc.decode("utf-8", errors="replace")
def _collect(code: str, budget_bytes: int, out_path: Path, seed: int = 0) -> int:
"""Stream parquet files (shuffled order) for one language until budget is met."""
repo, subdir = _source_for_code(code)
files = list(_list_parquet_files(repo, subdir))
random.Random(seed).shuffle(files)
got = 0
with open(out_path, "w", encoding="utf-8") as out:
for f in files:
if got >= budget_bytes:
break
try:
for doc in _iter_texts(repo, f):
for ln in _split_line(doc):
out.write(ln + "\n")
got += len(ln.encode("utf-8")) + 1
if got >= budget_bytes:
break
except Exception as exc:
print(f"[tokcorpus] WARN {repo}/{f} failed mid-stream: {exc}")
return got
def build_starved(total_bytes: float = 4e9, T: float = 100.0, seed: int = 0,
n_langs: int = N_STARVED_LANGS) -> Path:
"""T-temperature sample over ~419 FineWeb/FineWeb2 languages (ATLAS-scale replication)."""
universe = select_starved_languages(n_langs)
out_dir = ensure(TOK_CORPORA / "starved")
weights = {l: b ** (1.0 / T) for l, b in universe.items()}
z = sum(weights.values())
stats = {}
for i, code in enumerate(sorted(universe)):
budget = int(total_bytes * weights[code] / z)
out_path = out_dir / f"{code}.txt"
if out_path.exists() and out_path.stat().st_size >= 0.9 * budget:
stats[code] = {"budget": budget, "bytes": out_path.stat().st_size, "cached": True}
continue
got = _collect(code, budget, out_path, seed=seed + i)
stats[code] = {"budget": budget, "bytes": got}
print(f"[starved] {code}: {got/1e6:.1f}MB / budget {budget/1e6:.1f}MB "
f"({i+1}/{len(universe)})")
(out_dir / "stats.json").write_text(json.dumps(
{"total_bytes": total_bytes, "T": T, "n_langs": len(universe), "per_lang": stats},
indent=2))
return out_dir
def build_destarved(total_bytes: float = 4e9, seed: int = 0) -> Path:
"""5 study languages; byte budgets scaled by FLORES+ byte premium (content-uniform)."""
from ..byte_premium import load_premiums
premiums = load_premiums()
out_dir = ensure(TOK_CORPORA / "destarved")
z = sum(premiums[l] for l in LANGS)
stats = {}
for i, code in enumerate(LANGS):
budget = int(total_bytes * premiums[code] / z)
out_path = out_dir / f"{code}.txt"
if out_path.exists() and out_path.stat().st_size >= 0.9 * budget:
stats[code] = {"budget": budget, "bytes": out_path.stat().st_size, "cached": True}
continue
got = _collect(code, budget, out_path, seed=seed + i)
stats[code] = {"budget": budget, "bytes": got}
print(f"[destarved] {code}: {got/1e6:.1f}MB / budget {budget/1e6:.1f}MB")
(out_dir / "stats.json").write_text(json.dumps(
{"total_bytes": total_bytes, "premiums": premiums, "per_lang": stats}, indent=2))
return out_dir
def corpus_files(condition: str) -> list[Path]:
d = TOK_CORPORA / condition
files = sorted(d.glob("*.txt"))
if not files:
raise FileNotFoundError(f"no corpus at {d} - run `xscript tok-corpus --condition {condition}`")
return files
|