File size: 13,349 Bytes
26d5b81 | 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 | #!/usr/bin/env python3
"""
NeuroFlow 训练数据预处理工具
将 D:\语料\ 中的原始语料预处理为高效的 .tok1 二进制格式。
存放于 WSL 虚拟盘 (速度 ~200 MB/s vs HDD ~125 MB/s)。
用法:
python3 scripts/prepare_training_data.py \
--corpus D:/语料 \
--tokenizer configs/tokenizer_128k.json \
--output /home/user/neuroflow_data \
--max-seq-len 128 \
--max-samples 5000000
输出目录结构:
output/
train.tok1 # 训练数据
stats.json # 统计信息
"""
import argparse
import json
import os
import struct
import sys
import time
from pathlib import Path
# ─── BPE Tokenizer ───────────────────────────────────────────
class BPETokenizer:
"""与 C++ BPETokenizer 兼容的 Python 实现"""
def __init__(self, config_path: str):
with open(config_path, 'r', encoding='utf-8') as f:
config = json.load(f)
self.vocab = config.get('vocab', {})
self.id_to_token = {v: k for k, v in self.vocab.items()}
self.merges = config.get('merges', [])
self.vocab_size = len(self.vocab)
# Build merge ranks
self.merge_ranks = {}
for i, (a, b) in enumerate(self.merges):
self.merge_ranks[(a, b)] = i
# Special tokens
self.pad_id = 0
self.unk_id = 1
self.bos_id = 2
self.eos_id = 3
def apply_bpe(self, token: str) -> str:
"""Apply BPE merges to a single token using priority queue algorithm"""
if len(token) <= 1 or not self.merge_ranks:
return token
symbols = list(token)
n = len(symbols)
import heapq
# Build linked list
next_link = list(range(1, n)) + [None]
prev_link = [None] + list(range(0, n - 1))
pq = []
def push_pair(i):
j = next_link[i]
if j is None:
return
pair = (symbols[i], symbols[j])
rank = self.merge_ranks.get(pair)
if rank is not None:
heapq.heappush(pq, (rank, i))
for i in range(n):
push_pair(i)
while pq:
rank, i = heapq.heappop(pq)
j = next_link[i]
if j is None:
continue
pair = (symbols[i], symbols[j])
if self.merge_ranks.get(pair) != rank:
continue
# Merge
symbols[i] = symbols[i] + symbols[j]
k = next_link[j]
next_link[i] = k
if k is not None:
prev_link[k] = i
if prev_link[i] is not None:
push_pair(prev_link[i])
push_pair(i)
# Collect result
result = []
i = 0
while i is not None:
result.append(symbols[i])
i = next_link[i]
return ''.join(result)
def encode(self, text: str, max_len: int = 128) -> list[int]:
"""Encode text to token IDs"""
ids = [self.bos_id]
i = 0
while i < len(text) and len(ids) < max_len - 1:
# Handle UTF-8 multi-byte
char = text[i]
byte_len = 1
if ord(char) >= 0x80:
if ord(char) < 0xE0:
byte_len = 2
elif ord(char) < 0xF0:
byte_len = 3
else:
byte_len = 4
byte_seq = text[i:i + byte_len]
bpe_result = self.apply_bpe(byte_seq)
if bpe_result in self.vocab:
ids.append(self.vocab[bpe_result])
else:
for c in bpe_result:
ids.append(self.vocab.get(c, self.unk_id))
i += byte_len
ids.append(self.eos_id)
if len(ids) > max_len:
ids = ids[:max_len - 1] + [self.eos_id]
return ids
def decode(self, ids: list[int]) -> str:
"""Decode token IDs to text"""
result = []
for id in ids:
if id in (self.pad_id, self.unk_id, self.bos_id, self.eos_id):
continue
token = self.id_to_token.get(id, '')
if token:
result.append(token)
return ''.join(result)
# ─── TOK1 Format Writer ─────────────────────────────────────
# TOK1 Binary Format:
# Magic: "TOK1" (4 bytes)
# Version: uint16 (2 bytes)
# VocabSize: uint32 (4 bytes)
# MaxSeqLen: uint32 (4 bytes)
# NumSamples:uint32 (4 bytes)
# For each sample:
# SeqLen: uint16 (2 bytes)
# TokenIDs: uint32[SeqLen] (4 bytes each)
class TOK1Writer:
def __init__(self, path: str, vocab_size: int, max_seq_len: int):
self.f = open(path, 'wb')
self.f.write(b'TOK1')
self.f.write(struct.pack('<H', 1)) # version
self.f.write(struct.pack('<I', vocab_size))
self.f.write(struct.pack('<I', max_seq_len))
self.f.write(struct.pack('<I', 0)) # num_samples placeholder
self.count = 0
self.max_seq_len = max_seq_len
self._pos_count = self.f.tell()
def add(self, token_ids: list[int]):
# Truncate if needed
if len(token_ids) > self.max_seq_len:
token_ids = token_ids[:self.max_seq_len]
self.f.write(struct.pack('<H', len(token_ids)))
self.f.write(struct.pack(f'<{len(token_ids)}I', *token_ids))
self.count += 1
def close(self):
# Update num_samples
self.f.seek(self._pos_count - 4) # back to num_samples field
self.f.write(struct.pack('<I', self.count))
self.f.close()
return self.count
# ─── Corpus Scanner ─────────────────────────────────────────
def scan_corpus(corpus_root: str) -> list[Path]:
"""Scan corpus directory and return list of text files"""
root = Path(corpus_root)
if not root.exists():
print(f"Error: corpus root not found: {corpus_root}")
sys.exit(1)
supported_exts = {'.txt', '.json', '.jsonl', '.csv', '.tsv', '.md'}
files = []
for ext in supported_exts:
found = list(root.rglob(f'*{ext}'))
files.extend(found)
if found:
print(f" {ext}: {len(found)} files")
# Sort for reproducibility
files.sort()
return files
def extract_text_from_file(filepath: Path) -> list[str]:
"""Extract text content from various file formats"""
texts = []
ext = filepath.suffix.lower()
try:
if ext == '.txt' or ext == '.md':
with open(filepath, 'r', encoding='utf-8', errors='ignore') as f:
content = f.read()
# Split into paragraphs
paragraphs = [p.strip() for p in content.split('\n\n') if len(p.strip()) >= 10]
texts.extend(paragraphs)
elif ext == '.json':
with open(filepath, 'r', encoding='utf-8', errors='ignore') as f:
content = f.read()
# Try JSON array
try:
data = json.loads(content)
if isinstance(data, list):
for item in data:
if isinstance(item, dict):
for key in ('text', 'content', 'title', 'question', 'answer'):
if key in item and isinstance(item[key], str):
texts.append(item[key])
elif isinstance(data, dict):
for key in ('text', 'content', 'title', 'question', 'answer'):
if key in data and isinstance(data[key], str):
texts.append(data[key])
except json.JSONDecodeError:
# Fallback: treat as plain text
if len(content) >= 10:
texts.append(content)
elif ext == '.jsonl':
with open(filepath, 'r', encoding='utf-8', errors='ignore') as f:
for line in f:
line = line.strip()
if not line or line.startswith('#'):
continue
try:
item = json.loads(line)
for key in ('text', 'content', 'title', 'question', 'answer'):
if key in item and isinstance(item[key], str):
texts.append(item[key])
break
except json.JSONDecodeError:
pass
elif ext in ('.csv', '.tsv'):
delim = '\t' if ext == '.tsv' else ','
with open(filepath, 'r', encoding='utf-8', errors='ignore') as f:
for line in f:
line = line.strip()
if not line or line.startswith('#'):
continue
# Take longest field as text
fields = line.split(delim)
if fields:
longest = max(fields, key=len)
if len(longest) >= 10:
texts.append(longest)
except Exception as e:
pass # Skip problematic files
return texts
# ─── Main ────────────────────────────────────────────────────
def main():
parser = argparse.ArgumentParser(description='NeuroFlow Training Data Preparer')
parser.add_argument('--corpus', required=True, help='Corpus root directory')
parser.add_argument('--tokenizer', required=True, help='Tokenizer config path')
parser.add_argument('--output', required=True, help='Output directory')
parser.add_argument('--max-seq-len', type=int, default=128, help='Max sequence length')
parser.add_argument('--max-samples', type=int, default=5000000, help='Max total samples')
parser.add_argument('--min-text-len', type=int, default=10, help='Minimum text length')
args = parser.parse_args()
os.makedirs(args.output, exist_ok=True)
# Load tokenizer
print(f"\n{'='*60}")
print(f"NeuroFlow 训练数据预处理")
print(f"{'='*60}")
print(f"语料根目录: {args.corpus}")
print(f"输出目录: {args.output}")
print(f"最大序列: {args.max_seq_len}")
print(f"最大样本: {args.max_samples:,}")
print()
print("加载分词器...")
tok = BPETokenizer(args.tokenizer)
print(f" 词表大小: {tok.vocab_size}")
print(f" Merges: {len(tok.merges)}")
# Scan corpus
print("\n扫描语料...")
files = scan_corpus(args.corpus)
print(f"总计: {len(files):,} 文件")
# Process files
print(f"\n开始处理...")
t0 = time.time()
total_chars = 0
total_samples = 0
skipped = 0
train_path = os.path.join(args.output, 'train.tok1')
writer = TOK1Writer(train_path, tok.vocab_size, args.max_seq_len)
for i, filepath in enumerate(files):
texts = extract_text_from_file(filepath)
for text in texts:
if len(text) < args.min_text_len:
skipped += 1
continue
try:
ids = tok.encode(text, args.max_seq_len)
if len(ids) >= 4: # at least bos + 2 tokens + eos
writer.add(ids)
total_samples += 1
total_chars += len(text)
except Exception:
skipped += 1
continue
if total_samples >= args.max_samples:
break
if total_samples >= args.max_samples:
break
# Progress report
if (i + 1) % 1000 == 0:
elapsed = time.time() - t0
rate = total_samples / elapsed if elapsed > 0 else 0
print(f" [{i+1}/{len(files)}] "
f"samples={total_samples:,} "
f"chars={total_chars:,} "
f"rate={rate:.0f} samples/s "
f"elapsed={elapsed:.1f}s")
final_count = writer.close()
elapsed = time.time() - t0
# Stats
file_size = os.path.getsize(train_path)
print(f"\n{'='*60}")
print(f"处理完成!")
print(f"{'='*60}")
print(f" 样本数: {final_count:,}")
print(f" 字符数: {total_chars:,}")
print(f" 跳过: {skipped:,}")
print(f" 文件大小: {file_size / (1024**3):.2f} GB")
print(f" 耗时: {elapsed:.1f}s ({elapsed/60:.1f}min)")
print(f" 速率: {final_count/elapsed:.0f} samples/s")
print(f" 输出: {train_path}")
# Save stats
stats = {
'num_samples': final_count,
'total_chars': total_chars,
'skipped': skipped,
'file_size_bytes': file_size,
'vocab_size': tok.vocab_size,
'max_seq_len': args.max_seq_len,
'corpus_root': args.corpus,
'elapsed_seconds': elapsed,
}
stats_path = os.path.join(args.output, 'stats.json')
with open(stats_path, 'w') as f:
json.dump(stats, f, indent=2, ensure_ascii=False)
print(f" 统计: {stats_path}")
if __name__ == '__main__':
main()
|