File size: 9,746 Bytes
d2e4870 | 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 | """Train the study's tokenizers.
We pretrain models with TWO SentencePiece Unigram tokenizers only. The `bpe`
and `pa` flavors are trained purely as tokenizer-analysis comparators for the
fertility/allocation gate (`xscript tok-analyze`); no model run ever uses them.
MODEL-TRAINING tokenizers -- SentencePiece Unigram, character_coverage=0.999995,
byte fallback:
unigram_starved -- ATLAS-style replication arm: T=100 temperature mixture
over ~419 languages. Matches both ATLAS's ~uniform 420-
language mixture and the Unigram algorithm of the MADLAD-
400 lineage its tokenizer descends from.
unigram_destarved -- the intervention arm: our 5 study languages only, byte-
premium content-aligned (equal *content*, not bytes, per
language; see data/tokcorpus.py). Same algorithm as the
starved arm, so the starved-vs-destarved contrast isolates
vocabulary allocation rather than confounding it with the
tokenizer algorithm.
On the algorithm choice: Unigram is MADLAD-400's confirmed algorithm (its
released 256k *model* tokenizer is SentencePiece Unigram). ATLAS's 64k tokenizer
is a SEPARATE artifact -- trained by the MADLAD-400 authors (Kudugunta et al.)
on the same T=100 recipe -- whose algorithm ATLAS does not state in-text, though
Unigram is the natural inference from that lineage. Do not conflate ATLAS's 64k
with MADLAD's 256k; they are different tokenizers.
ANALYSIS-ONLY comparators -- trained for the gate, never used to pretrain:
bpe -- byte-level BPE (Whitespace + ByteLevel pre-tokenization) trained with
HuggingFace `tokenizers`' Rust `BpeTrainer`. Quantifies how much the
Unigram-vs-BPE algorithm choice alone moves fertility/allocation.
pa -- parity-aware byte-level BPE via swiss-ai/parity-aware-bpe's
`parity_aware_learn_bpe.py` (window variant, for ZH), fertility-
equalized over the 5-way-parallel FLORES+ dev set. Same byte-level
alphabet as `bpe`; the merge criterion (parity-balanced vs frequency)
is the only difference -> a clean upper bound on fertility
equalization. Destarved only (it balances a fixed dev-language set).
Uses the slow single-threaded reference trainer -- tolerable only
because its corpus is 5 languages, not 419.
Every flavor exposes exactly `VOCAB_SIZE` pieces with our four specials at ids
0..3, so packed token ids stay uint16 and every downstream module stays flavor-
agnostic. VOCAB_SIZE is overridable via XSCRIPT_VOCAB for the CPU smoke test.
"""
import json
import os
import subprocess
from pathlib import Path
from ..langs import tok_name
from ..paths import TOK_CORPORA, tokenizer_dir, ensure
from ..data.tokcorpus import corpus_files
VOCAB_SIZE = int(os.environ.get("XSCRIPT_VOCAB", "65536"))
SPECIALS = ["<unk>", "<bos>", "<eos>", "<pad>"] # ids 0..3 in every flavor
PA_REPO = "swiss-ai/parity-aware-bpe"
# --------------------------------------------------------------------------- #
# unigram (SentencePiece)
# --------------------------------------------------------------------------- #
def train_unigram(condition: str, seed: int = 42) -> Path:
import sentencepiece as spm
if hasattr(spm, "set_random_generator_seed"):
spm.set_random_generator_seed(seed) # not a TrainerSpec field in >=0.2
files = corpus_files(condition)
out = ensure(tokenizer_dir(tok_name("unigram", condition)))
spm.SentencePieceTrainer.train(
input=",".join(str(f) for f in files),
model_prefix=str(out / "sp"),
model_type="unigram",
vocab_size=VOCAB_SIZE,
character_coverage=0.999995,
byte_fallback=True,
unk_id=0, bos_id=1, eos_id=2, pad_id=3,
unk_piece="<unk>", bos_piece="<bos>", eos_piece="<eos>", pad_piece="<pad>",
input_sentence_size=10_000_000,
shuffle_input_sentence=True,
train_extremely_large_corpus=True,
remove_extra_whitespaces=False,
num_threads=max(1, (os.cpu_count() or 8) - 2),
)
_write_meta(out, "unigram", condition, files)
return out
# --------------------------------------------------------------------------- #
# byte-level BPE + parity-aware BPE (swiss-ai/parity-aware-bpe)
# --------------------------------------------------------------------------- #
def _n_merges() -> int:
# vocab = 4 specials + 256 byte-level base alphabet + merges
return VOCAB_SIZE - len(SPECIALS) - 256
def train_bpe(condition: str) -> Path:
from tokenizers import Tokenizer, models, pre_tokenizers, decoders, trainers
files = corpus_files(condition)
out = ensure(tokenizer_dir(tok_name("bpe", condition)))
tok = Tokenizer(models.BPE(unk_token=None, fuse_unk=False))
tok.pre_tokenizer = pre_tokenizers.Sequence(
[pre_tokenizers.Whitespace(), pre_tokenizers.ByteLevel(use_regex=False)])
tok.decoder = decoders.ByteLevel()
trainer = trainers.BpeTrainer(
vocab_size=VOCAB_SIZE,
special_tokens=SPECIALS, # ids 0..3, in order
initial_alphabet=pre_tokenizers.ByteLevel.alphabet(), # full 256 bytes
show_progress=True,
)
tok.train([str(f) for f in files], trainer)
tok.save(str(out / "tokenizer.json"))
_write_meta(out, "bpe", condition, files,
extra={"vocab_size_actual": tok.get_vocab_size(),
"source": "huggingface-tokenizers-bpe"})
return out
def train_pa(condition: str = "destarved", variant: str = "window") -> Path:
if condition != "destarved":
raise ValueError("parity-aware BPE is destarved-only (see langs.tok_conditions)")
inputs = corpus_files("destarved") # one file per study language
dev = _write_pa_dev(inputs) # aligned FLORES+ dev per lang
out = ensure(tokenizer_dir(tok_name("pa", condition)))
merges = out / "merges.raw.txt"
# parity-aware's multi-worker vocab builder is broken in the released
# version (pickle.load on a text-mode file), so force single-worker.
pa_workers = os.environ.get("XSCRIPT_PA_WORKERS", "1")
cmd = ["python", "-m", "parity_aware_bpe.parity_aware_learn_bpe",
"--variant", variant, "--symbols", str(_n_merges()),
"--num-workers", pa_workers, "--output", str(merges),
"--input", *[str(f) for f in inputs],
"--dev", *[str(f) for f in dev]]
_run(cmd)
_bytelevel_from_merges(merges, out, "pa", condition, inputs)
return out
def _write_pa_dev(inputs) -> list[Path]:
"""FLORES+ dev text per language, in the SAME order as `inputs` (stem=code)."""
from .. import flores
d = ensure(TOK_CORPORA / "pa_dev")
dev = []
for f in inputs:
code = f.stem
sents = list(flores.load(code, "dev").values())
p = d / f"{code}.dev.txt"
p.write_text("\n".join(sents) + "\n", encoding="utf-8")
dev.append(p)
return dev
def _bytelevel_from_merges(merges_path: Path, out: Path, flavor: str,
condition: str, corpus_files_used) -> None:
"""Merge rules -> HuggingFace byte-level BPE tokenizer, exactly VOCAB_SIZE."""
from tokenizers import Tokenizer, models, pre_tokenizers, decoders
lines = [l.strip() for l in merges_path.read_text(encoding="utf-8").splitlines()
if l.strip()]
if lines and lines[0].startswith("#version"):
lines = lines[1:]
vocab: dict[str, int] = {s: i for i, s in enumerate(SPECIALS)} # 0..3
for ch in pre_tokenizers.ByteLevel.alphabet(): # 256 bytes
vocab.setdefault(ch, len(vocab))
keep = max(0, VOCAB_SIZE - len(vocab)) # merges budget
merges: list[tuple[str, str]] = []
for line in lines:
if len(merges) >= keep:
break
a, b = line.split(" ")
if a not in vocab or b not in vocab: # order guarantees this won't hit
continue
merges.append((a, b))
vocab.setdefault(a + b, len(vocab))
tok = Tokenizer(models.BPE(vocab=vocab, merges=merges,
unk_token=None, fuse_unk=False))
# EXACT pre-tokenizer/decoder the repo trains and loads with (byte-level)
tok.pre_tokenizer = pre_tokenizers.Sequence(
[pre_tokenizers.Whitespace(), pre_tokenizers.ByteLevel(use_regex=False)])
tok.decoder = decoders.ByteLevel()
tok.save(str(out / "tokenizer.json"))
_write_meta(out, flavor, condition, corpus_files_used,
extra={"vocab_size_actual": tok.get_vocab_size(),
"n_merges": len(merges), "source": PA_REPO})
def _run(cmd, shell: bool = False) -> None:
print(f"[tok] $ {cmd if shell else ' '.join(cmd)}")
subprocess.run(cmd, shell=shell, check=True)
# --------------------------------------------------------------------------- #
def _write_meta(out: Path, flavor: str, condition: str, files, extra=None) -> None:
meta = {
"flavor": flavor,
"condition": condition,
"vocab_size": VOCAB_SIZE,
"specials": SPECIALS,
"corpus_files": [str(f) for f in files],
}
if extra:
meta.update(extra)
(out / "meta.json").write_text(json.dumps(meta, indent=2))
print(f"[tok] trained {flavor}_{condition} -> {out}")
def train(flavor: str, condition: str) -> Path:
if flavor == "unigram":
return train_unigram(condition)
if flavor == "bpe":
return train_bpe(condition)
if flavor == "pa":
return train_pa(condition)
raise ValueError(f"unknown flavor {flavor!r} (want unigram|bpe|pa)")
|