frankenstallm / data /tokenize_extra.py
pathcosmos's picture
feat: Add data pipeline scripts + phase reports (Tier 3 - reproducibility)
b3d361d verified
"""
data/tokenize_extra.py β€” λŒ€μš©λŸ‰ korean_extra/ 데이터셋 병렬 토큰화
HuggingFace datasets disk 포맷(arrow), parquet, jsonl λ“± μ„Έ κ°€μ§€ 포맷을
μžλ™ κ°μ§€ν•˜μ—¬ SentencePiece ν† ν¬λ‚˜μ΄μ €λ‘œ ν† ν°ν™”ν•˜κ³ , κ²°κ³Όλ₯Ό uint16 memmap
(.bin) 파일둜 μ €μž₯ν•œλ‹€. 881 GB μ΄μƒμ˜ λŒ€μš©λŸ‰ 데이터도 슀트리밍·청크 λ°©μ‹μœΌλ‘œ
μ²˜λ¦¬ν•œλ‹€.
좜λ ₯ 포맷은 data/dataset.py PackedDataset / TextDataset κ³Ό μ™„μ „νžˆ ν˜Έν™˜λ˜λŠ”
numpy uint16 ν”Œλž« 배열이닀.
μ‚¬μš© μ˜ˆμ‹œ:
# 단일 디렉토리
python data/tokenize_extra.py \
--input_dir data/korean_extra/fineweb2_edu_ko \
--output data/fineweb2_train.bin \
--num_proc 8
# korean_extra/ 전체 μ„œλΈŒλ””λ ‰ν† λ¦¬ 일괄 처리
python data/tokenize_extra.py \
--input_dir data/korean_extra \
--auto_scan \
--output_dir data \
--num_proc 8
# 곡개 검증
python -c "
import numpy as np
d = np.memmap('data/fineweb2_train.bin', dtype='uint16', mode='r')
print(f'총 토큰: {len(d):,}')
"
"""
from __future__ import annotations
import argparse
import json
import math
import multiprocessing as mp
import os
import struct
import sys
import time
from pathlib import Path
from typing import Generator, Iterable, Iterator
import numpy as np
from tqdm import tqdm
# ---------------------------------------------------------------------------
# SentencePiece μž„ν¬νŠΈ (선택적 β€” μ—†μœΌλ©΄ 였λ₯˜ λ©”μ‹œμ§€ 좜λ ₯ ν›„ μ’…λ£Œ)
# ---------------------------------------------------------------------------
try:
import sentencepiece as spm
except ImportError:
print(
"ERROR: sentencepiece νŒ¨ν‚€μ§€κ°€ μ„€μΉ˜λ˜μ–΄ μžˆμ§€ μ•ŠμŠ΅λ‹ˆλ‹€.\n"
" pip install sentencepiece 둜 μ„€μΉ˜ ν›„ μž¬μ‹€ν–‰ν•˜μ„Έμš”.",
file=sys.stderr,
)
sys.exit(1)
# ---------------------------------------------------------------------------
# datasets μž„ν¬νŠΈ
# ---------------------------------------------------------------------------
try:
import datasets as hf_datasets
except ImportError:
print(
"ERROR: datasets νŒ¨ν‚€μ§€κ°€ μ„€μΉ˜λ˜μ–΄ μžˆμ§€ μ•ŠμŠ΅λ‹ˆλ‹€.\n"
" pip install datasets 둜 μ„€μΉ˜ ν›„ μž¬μ‹€ν–‰ν•˜μ„Έμš”.",
file=sys.stderr,
)
sys.exit(1)
# ===========================================================================
# μƒμˆ˜
# ===========================================================================
UINT16_MAX = 65535 # uint16 μ˜€λ²„ν”Œλ‘œ 경계
MIN_TOKENS = 100 # μ΅œμ†Œ 토큰 수 (미만이면 버림)
MAX_TOKENS = 32_768 # μ΅œλŒ€ 토큰 수 (μ΄ˆκ³ΌλΆ„μ€ 버림)
HANGUL_RE_THRESHOLD = 0.10 # ν•œκΈ€ λΉ„μœ¨ μ΅œμ†Œ κΈ°μ€€ (이 미만이고 ν•œκΈ€ μ•„λ‹Œ 경우 버림)
CHUNK_TOKENS = 500_000 # memmap 청크 λ‹¨μœ„ (tokens)
EOS_TOKEN_PLACEHOLDER = 1 # EOS id β€” SP κΈ°λ³Έκ°’, μ‹€μ œ idλŠ” λͺ¨λΈμ—μ„œ 읽음
# ---------------------------------------------------------------------------
# ν•œκΈ€ λΉ„μœ¨ ν•„ν„°
# ---------------------------------------------------------------------------
# ord λ²”μœ„: κ°€(AC00) ~ 힣(D7A3), γ„±(3131) ~ γ…£(3163)
_HANGUL_START = 0xAC00
_HANGUL_END = 0xD7A3
def _has_enough_korean_or_english(text: str) -> bool:
"""
ν•œκΈ€ 문자 λΉ„μœ¨μ΄ HANGUL_RE_THRESHOLD μ΄μƒμ΄κ±°λ‚˜,
ASCII μ•ŒνŒŒλ²³ λΉ„μœ¨μ΄ 0.3 이상이면 True λ°˜ν™˜.
λ‘˜ λ‹€ μ•„λ‹Œ 경우 False (쀑ꡭ어, μΌλ³Έμ–΄λ§Œ μžˆλŠ” λ“±).
"""
if not text:
return False
total = len(text)
hangul_cnt = sum(1 for ch in text if _HANGUL_START <= ord(ch) <= _HANGUL_END)
if hangul_cnt / total >= HANGUL_RE_THRESHOLD:
return True
ascii_alpha = sum(1 for ch in text if ch.isascii() and ch.isalpha())
if ascii_alpha / total >= 0.30:
return True
return False
# ===========================================================================
# ν† ν¬λ‚˜μ΄μ € 래퍼 (ν”„λ‘œμ„ΈμŠ€ κ°„ 곡유 λΆˆκ°€ β€” 각 μ›Œμ»€μ—μ„œ reload)
# ===========================================================================
class SPTokenizer:
"""SentencePiece λͺ¨λΈμ„ wrappingν•œ κ°„λ‹¨ν•œ ν† ν¬λ‚˜μ΄μ €."""
def __init__(self, model_path: str) -> None:
self._model_path = model_path
self._sp: spm.SentencePieceProcessor | None = None
# ν”„λ‘œμ„ΈμŠ€ fork ν›„ _spκ°€ None인 경우 lazy load
def _ensure_loaded(self) -> None:
if self._sp is None:
sp = spm.SentencePieceProcessor()
sp.Load(self._model_path)
self._sp = sp
@property
def eos_id(self) -> int:
self._ensure_loaded()
return self._sp.eos_id()
@property
def vocab_size(self) -> int:
self._ensure_loaded()
return self._sp.GetPieceSize()
def encode(self, text: str) -> list[int]:
self._ensure_loaded()
return self._sp.EncodeAsIds(text)
# ===========================================================================
# 포맷 감지 & μ΄ν„°λ ˆμ΄ν„°
# ===========================================================================
def _detect_format(input_dir: Path) -> str:
"""
디렉토리 λ‚΄μš©μ„ 보고 포맷을 μžλ™ κ°μ§€ν•œλ‹€.
λ°˜ν™˜κ°’:
"hf_arrow" β€” HuggingFace datasets disk 포맷 (dataset_info.json 쑴재)
"parquet" β€” .parquet 파일이 있음
"jsonl" β€” .jsonl λ˜λŠ” .json 파일이 있음
"unknown" β€” μ•Œ 수 μ—†μŒ
"""
if not input_dir.is_dir():
raise NotADirectoryError(f"μž…λ ₯ κ²½λ‘œκ°€ 디렉토리가 μ•„λ‹™λ‹ˆλ‹€: {input_dir}")
# HF arrow 포맷 νŒλ³„ β€” dataset_info.json λ˜λŠ” state.json이 있으면 HF 포맷
if (input_dir / "dataset_info.json").exists():
return "hf_arrow"
if (input_dir / "state.json").exists():
return "hf_arrow"
# μ„œλΈŒ 디렉토리 μ•ˆμ— dataset_info.json이 μžˆλŠ” 경우 (split 포함)
for child in input_dir.iterdir():
if child.is_dir() and (child / "dataset_info.json").exists():
return "hf_arrow"
# parquet 파일 확인
parquets = list(input_dir.rglob("*.parquet"))
if parquets:
return "parquet"
# jsonl / json 파일 확인
jsonls = list(input_dir.rglob("*.jsonl")) + list(input_dir.rglob("*.json"))
if jsonls:
return "jsonl"
return "unknown"
def _iter_hf_arrow(
input_dir: Path,
text_col: str,
num_proc: int,
) -> Iterator[str]:
"""HuggingFace datasets disk ν¬λ§·μ—μ„œ ν…μŠ€νŠΈλ₯Ό μŠ€νŠΈλ¦¬λ°ν•œλ‹€."""
print(f" [포맷] HuggingFace arrow (disk): {input_dir}")
try:
ds = hf_datasets.load_from_disk(str(input_dir))
except Exception as exc:
# DatasetDict일 수 있음 β€” 'train' split μ‹œλ„
try:
ds_dict = hf_datasets.load_from_disk(str(input_dir))
if isinstance(ds_dict, hf_datasets.DatasetDict):
splits = list(ds_dict.keys())
print(f" DatasetDict 감지. splits={splits}, 'train' split μ‚¬μš©.")
ds = ds_dict.get("train", ds_dict[splits[0]])
else:
raise exc
except Exception:
raise RuntimeError(
f"HF arrow 포맷 λ‘œλ“œ μ‹€νŒ¨: {input_dir}\n원인: {exc}"
) from exc
# μ‹€μ œ ν…μŠ€νŠΈ 컬럼 이름 μΆ”μ •
col = _resolve_text_col(list(ds.column_names), text_col)
print(f" ν…μŠ€νŠΈ 컬럼: '{col}', 총 ν–‰ 수: {len(ds):,}")
for row in ds:
yield row[col]
def _iter_parquet(input_dir: Path, text_col: str) -> Iterator[str]:
"""parquet νŒŒμΌμ—μ„œ ν…μŠ€νŠΈλ₯Ό μŠ€νŠΈλ¦¬λ°ν•œλ‹€."""
try:
import pyarrow.parquet as pq # type: ignore
except ImportError:
# datasets둜 fallback
print(" [κ²½κ³ ] pyarrow λ―Έμ„€μΉ˜, datasets둜 parquet λ‘œλ“œ μ‹œλ„...")
files = sorted(input_dir.rglob("*.parquet"))
print(f" [포맷] parquet ({len(files)} 파일): {input_dir}")
ds = hf_datasets.load_dataset(
"parquet",
data_files={"train": [str(f) for f in files]},
split="train",
streaming=True,
)
col = _resolve_text_col(list(ds.column_names), text_col)
print(f" ν…μŠ€νŠΈ 컬럼: '{col}'")
for row in ds:
yield row[col]
return
files = sorted(input_dir.rglob("*.parquet"))
print(f" [포맷] parquet ({len(files)} 파일): {input_dir}")
for fpath in files:
pf = pq.ParquetFile(str(fpath))
cols = pf.schema_arrow.names
col = _resolve_text_col(cols, text_col)
for batch in pf.iter_batches(batch_size=1000, columns=[col]):
for val in batch.column(col):
yield val.as_py() or ""
def _iter_jsonl(input_dir: Path, text_col: str) -> Iterator[str]:
"""jsonl / json νŒŒμΌμ—μ„œ ν…μŠ€νŠΈλ₯Ό μŠ€νŠΈλ¦¬λ°ν•œλ‹€."""
files = sorted(input_dir.rglob("*.jsonl")) + sorted(input_dir.rglob("*.json"))
# json 파일 쀑 jsonl이 μ•„λ‹Œ 것 제거 (파일 μžμ²΄κ°€ dict인 경우)
print(f" [포맷] jsonl ({len(files)} 파일): {input_dir}")
for fpath in files:
try:
with open(fpath, "r", encoding="utf-8", errors="replace") as fh:
for line in fh:
line = line.strip()
if not line:
continue
try:
obj = json.loads(line)
except json.JSONDecodeError:
continue
if isinstance(obj, str):
yield obj
elif isinstance(obj, dict):
text = (
obj.get(text_col)
or obj.get("text")
or obj.get("content")
or obj.get("document")
or ""
)
yield str(text)
except Exception as exc:
print(f" [κ²½κ³ ] 파일 읽기 μ‹€νŒ¨: {fpath} β€” {exc}", file=sys.stderr)
def _resolve_text_col(columns: list[str], preferred: str) -> str:
"""
μ§€μ •λœ 컬럼이 없을 경우, 일반적인 ν…μŠ€νŠΈ 컬럼 이름을 μˆœμ„œλŒ€λ‘œ νƒμƒ‰ν•œλ‹€.
"""
if preferred in columns:
return preferred
for candidate in ("text", "content", "document", "body", "passage"):
if candidate in columns:
print(
f" [INFO] 컬럼 '{preferred}' 미쑴재 β†’ '{candidate}' μ‚¬μš©. "
f"(전체 컬럼: {columns[:10]})"
)
return candidate
# λ§ˆμ§€λ§‰ μˆ˜λ‹¨: 첫 번째 λ¬Έμžμ—΄ 컬럼
print(
f" [κ²½κ³ ] ν…μŠ€νŠΈ μ»¬λŸΌμ„ μ°Ύμ§€ λͺ»ν•¨. 첫 번째 컬럼 '{columns[0]}' μ‚¬μš©.",
file=sys.stderr,
)
return columns[0]
def get_text_iterator(
input_dir: Path,
text_col: str,
num_proc: int,
) -> tuple[str, Iterator[str]]:
"""
포맷을 μžλ™ κ°μ§€ν•˜κ³  μ•Œλ§žμ€ ν…μŠ€νŠΈ μ΄ν„°λ ˆμ΄ν„°λ₯Ό λ°˜ν™˜ν•œλ‹€.
Returns:
(fmt, iterator) fmt은 κ°μ§€λœ 포맷 λ¬Έμžμ—΄
"""
fmt = _detect_format(input_dir)
if fmt == "hf_arrow":
return fmt, _iter_hf_arrow(input_dir, text_col, num_proc)
elif fmt == "parquet":
return fmt, _iter_parquet(input_dir, text_col)
elif fmt == "jsonl":
return fmt, _iter_jsonl(input_dir, text_col)
else:
raise RuntimeError(
f"μ§€μ›ν•˜μ§€ μ•ŠλŠ” ν¬λ§·μ΄κ±°λ‚˜ 인식할 수 μ—†μŠ΅λ‹ˆλ‹€: {input_dir}\n"
f"지원 포맷: HuggingFace arrow, parquet, jsonl"
)
# ===========================================================================
# 단일 ν”„λ‘œμ„ΈμŠ€ 토큰화 μ›Œμ»€ (multiprocessing.Poolμ—μ„œ 호좜)
# ===========================================================================
# μ „μ—­ ν† ν¬λ‚˜μ΄μ € β€” 각 μ›Œμ»€ ν”„λ‘œμ„ΈμŠ€μ—μ„œ ν•œ 번만 μ΄ˆκΈ°ν™”
_g_sp: SPTokenizer | None = None
_g_model_path: str = ""
def _worker_init(model_path: str) -> None:
"""μ›Œμ»€ μ΄ˆκΈ°ν™” ν•¨μˆ˜: SentencePiece λͺ¨λΈ λ‘œλ“œ."""
global _g_sp, _g_model_path
_g_model_path = model_path
_g_sp = SPTokenizer(model_path)
_g_sp._ensure_loaded()
def _worker_tokenize_batch(texts: list[str]) -> list[list[int]]:
"""
ν…μŠ€νŠΈ 배치λ₯Ό ν† ν°ν™”ν•˜κ³  ν’ˆμ§ˆ ν•„ν„°λ₯Ό μ μš©ν•œλ‹€.
λ°˜ν™˜κ°’: μœ νš¨ν•œ 토큰 리슀트 λͺ©λ‘ (ν•„ν„° ν†΅κ³Όν•œ κ²ƒλ§Œ)
"""
global _g_sp
results: list[list[int]] = []
for text in texts:
if not text or not isinstance(text, str):
continue
# ν’ˆμ§ˆ ν•„ν„°: μ–Έμ–΄
if not _has_enough_korean_or_english(text):
continue
try:
ids = _g_sp.encode(text)
except Exception:
continue
# 길이 ν•„ν„°
if len(ids) < MIN_TOKENS:
continue
if len(ids) > MAX_TOKENS:
ids = ids[:MAX_TOKENS]
results.append(ids)
return results
# ===========================================================================
# memmap 청크 기반 기둝기
# ===========================================================================
class MemmapWriter:
"""
uint16 numpy memmap νŒŒμΌμ— 토큰을 청크 λ‹¨μœ„λ‘œ κΈ°λ‘ν•˜λŠ” 래퍼.
μ΄ˆκΈ°μ— μž‘μ€ 크기둜 μƒμ„±ν•˜κ³ , ν•„μš”ν•  λ•Œ resizeν•œλ‹€.
μ΅œμ’…μ μœΌλ‘œ μ‹€μ œ 기둝된 크기둜 truncateν•˜μ—¬ μ €μž₯ν•œλ‹€.
"""
def __init__(self, path: Path, initial_size: int = CHUNK_TOKENS) -> None:
self.path = path
path.parent.mkdir(parents=True, exist_ok=True)
self._alloc = max(initial_size, CHUNK_TOKENS)
self._mm = np.memmap(
str(path), dtype="uint16", mode="w+", shape=(self._alloc,)
)
self._pos = 0
def write(self, tokens: Iterable[int]) -> int:
"""tokensλ₯Ό κΈ°λ‘ν•˜κ³  기둝된 토큰 수λ₯Ό λ°˜ν™˜ν•œλ‹€."""
arr = np.asarray(list(tokens), dtype=np.uint16)
n = len(arr)
if n == 0:
return 0
needed = self._pos + n
if needed > self._alloc:
# 두 λ°° λ˜λŠ” ν•„μš”ν•œ 크기 쀑 큰 κ°’μœΌλ‘œ ν™•μž₯
new_alloc = max(self._alloc * 2, needed + CHUNK_TOKENS)
self._mm.flush()
del self._mm
self._alloc = new_alloc
self._mm = np.memmap(
str(self.path), dtype="uint16", mode="r+", shape=(self._alloc,)
)
self._mm[self._pos : self._pos + n] = arr
self._pos += n
return n
def finalize(self) -> int:
"""기둝된 μ‹€μ œ 크기둜 νŒŒμΌμ„ truncateν•˜κ³  λ‹«λŠ”λ‹€. 총 토큰 수λ₯Ό λ°˜ν™˜ν•œλ‹€."""
self._mm.flush()
del self._mm
# μ‹€μ œ 기둝된 크기둜 truncate
final_bytes = self._pos * 2 # uint16 = 2 bytes
with open(str(self.path), "r+b") as fh:
fh.truncate(final_bytes)
return self._pos
# ===========================================================================
# 핡심 토큰화 νŒŒμ΄ν”„λΌμΈ
# ===========================================================================
def tokenize_directory(
input_dir: Path,
output_path: Path,
tokenizer_path: str,
text_col: str = "text",
num_proc: int = 8,
batch_size: int = 512,
eos_between_docs: bool = True,
val_split: float = 0.002,
seed: int = 42,
) -> dict:
"""
단일 디렉토리λ₯Ό ν† ν°ν™”ν•˜μ—¬ .bin 파일(λ“€)둜 μ €μž₯ν•œλ‹€.
Args:
input_dir: μž…λ ₯ 디렉토리 (포맷 μžλ™ 감지)
output_path: 좜λ ₯ .bin 파일 경둜 (ν›ˆλ ¨ μ…‹)
tokenizer_path: SentencePiece .model 파일 경둜
text_col: ν…μŠ€νŠΈ 컬럼 이름 (arrow/parquetμ—μ„œ μ‚¬μš©)
num_proc: 병렬 μ›Œμ»€ 수
batch_size: μ›Œμ»€λ‹Ή 배치 크기
eos_between_docs: λ¬Έμ„œ 사이에 EOS 토큰 μ‚½μž… μ—¬λΆ€
val_split: 검증 뢄리 λΉ„μœ¨ (0 이면 val 파일 생성 μ•ˆ 함)
seed: μž¬ν˜„μ„± μ‹œλ“œ
Returns:
톡계 dict (total_tokens, train_tokens, val_tokens, skipped, elapsed_s)
"""
t_start = time.time()
# ─── ν† ν¬λ‚˜μ΄μ € λ‘œλ“œ (메인 ν”„λ‘œμ„ΈμŠ€: EOS id 확인) ─────────────────────
sp_main = SPTokenizer(tokenizer_path)
eos_id = sp_main.eos_id
vocab_size = sp_main.vocab_size
print(f" ν† ν¬λ‚˜μ΄μ €: {tokenizer_path}")
print(f" vocab_size={vocab_size:,}, eos_id={eos_id}")
if vocab_size > UINT16_MAX:
print(
f" [κ²½κ³ ] vocab_size({vocab_size}) > {UINT16_MAX} "
f"β€” uint16 μ˜€λ²„ν”Œλ‘œ κ°€λŠ₯. 65535 μ΄ν•˜ id만 μ•ˆμ „.",
file=sys.stderr,
)
# ─── 포맷 감지 & μ΄ν„°λ ˆμ΄ν„° 생성 ─────────────────────────────────────
fmt, text_iter = get_text_iterator(input_dir, text_col, num_proc)
print(f" 포맷: {fmt}")
# ─── 좜λ ₯ 경둜 μ„€μ • ────────────────────────────────────────────────────
train_path = output_path
val_path: Path | None = None
if val_split > 0:
stem = output_path.stem
if "train" in stem:
val_path = output_path.parent / output_path.name.replace("train", "val")
else:
val_path = output_path.with_name(stem + "_val" + output_path.suffix)
print(f" 좜λ ₯(train): {train_path}")
if val_path:
print(f" 좜λ ₯(val): {val_path}")
# ─── memmap 기둝기 μ΄ˆκΈ°ν™” ─────────────────────────────────────────────
writer = MemmapWriter(train_path)
val_writer: MemmapWriter | None = MemmapWriter(val_path) if val_path else None
# ─── multiprocessing Pool 생성 ────────────────────────────────────────
pool = mp.Pool(
processes=num_proc,
initializer=_worker_init,
initargs=(tokenizer_path,),
)
total_docs = 0
skipped = 0
total_toks = 0
# numpy rng for deterministic val split
rng = np.random.default_rng(seed)
def _submit_batch(batch_texts: list[str]) -> None:
nonlocal total_docs, skipped, total_toks
# 동기 map (배치 λ‹¨μœ„, μ›Œμ»€λ³„ μ„œλΈŒλ°°μΉ˜λ‘œ λΆ„ν• )
sub_size = max(1, len(batch_texts) // num_proc)
sub_batches = [
batch_texts[i : i + sub_size]
for i in range(0, len(batch_texts), sub_size)
]
results_list = pool.map(_worker_tokenize_batch, sub_batches)
for results in results_list:
for ids in results:
total_docs += 1
n = len(ids)
total_toks += n
# EOS 토큰 μ‚½μž…
if eos_between_docs:
ids_out = ids + [eos_id]
else:
ids_out = ids
# val split: λ¬΄μž‘μœ„λ‘œ val_split λΉ„μœ¨λ§ŒνΌ val 파일둜
if val_writer is not None and rng.random() < val_split:
val_writer.write(ids_out)
else:
writer.write(ids_out)
skipped_in_batch = sum(1 for _ in results) - len(results)
# ─── 배치 μˆ˜μ§‘ & tqdm μ§„ν–‰λ₯  ─────────────────────────────────────────
batch_buf: list[str] = []
pbar = tqdm(desc=f"토큰화 [{input_dir.name}]", unit="doc", dynamic_ncols=True)
for text in text_iter:
batch_buf.append(text)
if len(batch_buf) >= batch_size * num_proc:
_submit_batch(batch_buf)
pbar.update(len(batch_buf))
pbar.set_postfix(
tokens=f"{total_toks:,}",
docs=f"{total_docs:,}",
refresh=False,
)
batch_buf = []
# λ§ˆμ§€λ§‰ μž”μ—¬ 배치 처리
if batch_buf:
_submit_batch(batch_buf)
pbar.update(len(batch_buf))
pbar.close()
pool.close()
pool.join()
# ─── 파일 마무리 ──────────────────────────────────────────────────────
train_tokens = writer.finalize()
val_tokens = val_writer.finalize() if val_writer else 0
elapsed = time.time() - t_start
total_toks_with_eos = train_tokens + val_tokens
print()
print(f" μ™„λ£Œ: {elapsed:.1f}초")
print(f" 처리 λ¬Έμ„œ: {total_docs:,}")
print(f" 총 토큰(EOS 포함): {total_toks_with_eos:,}")
print(f" train: {train_tokens:,} ({train_tokens*2/1e9:.2f} GB)")
if val_tokens:
print(f" val: {val_tokens:,} ({val_tokens*2/1e9:.2f} GB)")
throughput = total_toks_with_eos / elapsed if elapsed > 0 else 0
print(f" 처리율: {throughput/1e6:.2f} M token/s")
return {
"total_docs" : total_docs,
"total_tokens" : total_toks_with_eos,
"train_tokens" : train_tokens,
"val_tokens" : val_tokens,
"elapsed_s" : elapsed,
"train_path" : str(train_path),
"val_path" : str(val_path) if val_path else None,
}
# ===========================================================================
# μ„œλΈŒλ””λ ‰ν† λ¦¬ μžλ™ μŠ€μΊ” λͺ¨λ“œ
# ===========================================================================
def auto_scan_and_tokenize(
root_dir: Path,
output_dir: Path,
tokenizer_path: str,
text_col: str,
num_proc: int,
batch_size: int,
val_split: float,
seed: int,
) -> list[dict]:
"""
root_dir 의 직접 μžμ‹ 디렉토리λ₯Ό μŠ€μΊ”ν•˜μ—¬ 각각 ν† ν°ν™”ν•œλ‹€.
각 μ„œλΈŒλ””λ ‰ν† λ¦¬μ— λŒ€ν•΄:
output_dir/korean_extra_{subdir_name}_train.bin 을 μƒμ„±ν•œλ‹€.
"""
children = sorted(p for p in root_dir.iterdir() if p.is_dir())
if not children:
raise RuntimeError(f"μ„œλΈŒλ””λ ‰ν† λ¦¬κ°€ μ—†μŠ΅λ‹ˆλ‹€: {root_dir}")
print(f"μžλ™ μŠ€μΊ”: {len(children)}개 μ„œλΈŒλ””λ ‰ν† λ¦¬ 발견")
for ch in children:
print(f" - {ch.name}")
print()
all_stats = []
for child in children:
print("=" * 60)
print(f"처리 쀑: {child}")
print("=" * 60)
safe_name = child.name.replace("/", "_").replace(" ", "_")
out_name = f"korean_extra_{safe_name}_train.bin"
out_path = output_dir / out_name
try:
stats = tokenize_directory(
input_dir = child,
output_path = out_path,
tokenizer_path = tokenizer_path,
text_col = text_col,
num_proc = num_proc,
batch_size = batch_size,
val_split = val_split,
seed = seed,
)
stats["source"] = child.name
all_stats.append(stats)
except Exception as exc:
print(f" [였λ₯˜] {child.name} 처리 μ‹€νŒ¨: {exc}", file=sys.stderr)
all_stats.append({"source": child.name, "error": str(exc)})
print()
return all_stats
# ===========================================================================
# CLI
# ===========================================================================
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description=(
"korean_extra/ λŒ€μš©λŸ‰ 데이터셋을 병렬 ν† ν°ν™”ν•˜μ—¬ uint16 memmap(.bin) 둜 μ €μž₯. "
"HuggingFace arrow, parquet, jsonl 포맷 μžλ™ 감지."
),
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
# μž…λ ₯
parser.add_argument(
"--input_dir",
required=True,
help="토큰화할 디렉토리 경둜. --auto_scan μ‹œμ—λŠ” 루트 디렉토리.",
)
parser.add_argument(
"--auto_scan",
action="store_true",
help=(
"input_dir 의 직접 μžμ‹ 디렉토리λ₯Ό λͺ¨λ‘ 순차 처리. "
"이 경우 --output_dir 을 μ§€μ •ν•΄μ•Ό 함."
),
)
parser.add_argument(
"--text_col",
default="text",
help="ν…μŠ€νŠΈ 컬럼 이름 (arrow/parquet/jsonl). μžλ™ μΆ”μ • κ°€λŠ₯.",
)
# 좜λ ₯
out_group = parser.add_mutually_exclusive_group()
out_group.add_argument(
"--output",
default=None,
help="좜λ ₯ .bin 파일 경둜 (단일 디렉토리 처리 μ‹œ μ‚¬μš©).",
)
out_group.add_argument(
"--output_dir",
default=None,
help="좜λ ₯ .bin νŒŒμΌλ“€μ„ μ €μž₯ν•  디렉토리 (--auto_scan μ‹œ μ‚¬μš©).",
)
# ν† ν¬λ‚˜μ΄μ €
parser.add_argument(
"--tokenizer",
default=(
"/PROJECT/0325120031_A/ghong/taketimes/llm-bang"
"/tokenizer/korean_64k.model"
),
help="SentencePiece .model 파일 경둜.",
)
# 처리 μ˜΅μ…˜
parser.add_argument(
"--num_proc",
type=int,
default=8,
help="병렬 μ›Œμ»€ 수 (multiprocessing.Pool).",
)
parser.add_argument(
"--batch_size",
type=int,
default=512,
help="μ›Œμ»€λ‹Ή 배치 크기 (λ¬Έμ„œ 수).",
)
parser.add_argument(
"--val_split",
type=float,
default=0.002,
help="검증 뢄리 λΉ„μœ¨ (0.0 이면 val 파일 미생성).",
)
parser.add_argument(
"--seed",
type=int,
default=42,
help="μž¬ν˜„μ„± μ‹œλ“œ.",
)
parser.add_argument(
"--no_eos",
action="store_true",
help="λ¬Έμ„œ 사이에 EOS 토큰을 μ‚½μž…ν•˜μ§€ μ•ŠλŠ”λ‹€.",
)
args = parser.parse_args()
# 검증
if not args.auto_scan and args.output is None:
# μžλ™ 좜λ ₯ 경둜 생성
input_name = Path(args.input_dir).name
args.output = str(
Path(args.input_dir).parent.parent
/ f"korean_extra_{input_name}_train.bin"
)
print(f"[INFO] --output λ―Έμ§€μ • β†’ μžλ™ μ„€μ •: {args.output}")
if args.auto_scan and args.output_dir is None:
parser.error("--auto_scan μ‚¬μš© μ‹œ --output_dir 을 μ§€μ •ν•΄μ•Ό ν•©λ‹ˆλ‹€.")
return args
def main() -> None:
args = parse_args()
tokenizer_path = args.tokenizer
if not Path(tokenizer_path).exists():
# fallback: μƒλŒ€κ²½λ‘œ μ‹œλ„
fallback = Path(
"/PROJECT/0325120031_A/ghong/taketimes/llm-bang"
"/tokenizer/korean_64k.model"
)
if fallback.exists():
tokenizer_path = str(fallback)
else:
print(
f"ERROR: ν† ν¬λ‚˜μ΄μ € νŒŒμΌμ„ 찾을 수 μ—†μŠ΅λ‹ˆλ‹€: {tokenizer_path}",
file=sys.stderr,
)
sys.exit(1)
print("=" * 60)
print(" LLM-Bang tokenize_extra.py")
print("=" * 60)
print(f" μž…λ ₯: {args.input_dir}")
print(f" ν† ν¬λ‚˜μ΄μ €: {tokenizer_path}")
print(f" num_proc: {args.num_proc}")
print(f" batch_size: {args.batch_size}")
print(f" val_split: {args.val_split}")
print(f" seed: {args.seed}")
print(f" eos: {not args.no_eos}")
print()
if args.auto_scan:
stats_list = auto_scan_and_tokenize(
root_dir = Path(args.input_dir),
output_dir = Path(args.output_dir),
tokenizer_path = tokenizer_path,
text_col = args.text_col,
num_proc = args.num_proc,
batch_size = args.batch_size,
val_split = args.val_split,
seed = args.seed,
)
print("=" * 60)
print(" 전체 μš”μ•½")
print("=" * 60)
grand_train = 0
grand_val = 0
for s in stats_list:
if "error" in s:
print(f" {s['source']:40s} ERROR: {s['error']}")
else:
t = s.get("train_tokens", 0)
v = s.get("val_tokens", 0)
grand_train += t
grand_val += v
print(
f" {s['source']:40s} "
f"train={t:>14,} val={v:>12,} "
f"({s['elapsed_s']:.0f}s)"
)
print("-" * 60)
print(
f" {'합계':40s} "
f"train={grand_train:>14,} val={grand_val:>12,}"
)
print(
f"\n 총 토큰: {grand_train + grand_val:,} "
f"({(grand_train + grand_val) * 2 / 1e9:.2f} GB)"
)
else:
stats = tokenize_directory(
input_dir = Path(args.input_dir),
output_path = Path(args.output),
tokenizer_path = tokenizer_path,
text_col = args.text_col,
num_proc = args.num_proc,
batch_size = args.batch_size,
eos_between_docs = not args.no_eos,
val_split = args.val_split,
seed = args.seed,
)
print()
print("=" * 60)
print(" κ²°κ³Ό μš”μ•½")
print("=" * 60)
print(f" train .bin : {stats['train_path']}")
if stats.get("val_path"):
print(f" val .bin : {stats['val_path']}")
print(f" train 토큰 : {stats['train_tokens']:,}")
print(f" val 토큰 : {stats['val_tokens']:,}")
print(f" 처리 λ¬Έμ„œ : {stats['total_docs']:,}")
print(f" μ†Œμš” μ‹œκ°„ : {stats['elapsed_s']:.1f}초")
# 검증: memmap λ‘œλ“œ ν…ŒμŠ€νŠΈ
print()
print(" [검증] memmap λ‘œλ“œ ν…ŒμŠ€νŠΈ...")
try:
d = np.memmap(stats["train_path"], dtype="uint16", mode="r")
print(f" memmap shape: {d.shape} dtype: {d.dtype}")
print(f" 첫 10 토큰: {d[:10].tolist()}")
except Exception as exc:
print(f" [κ²½κ³ ] memmap λ‘œλ“œ μ‹€νŒ¨: {exc}", file=sys.stderr)
if __name__ == "__main__":
main()