File size: 7,544 Bytes
187a491 | 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 | """
BoccaccioAI - BPE Tokenizer Training Script
Trains a byte-level BPE tokenizer on a subset of CulturaX Italian data
using the HuggingFace tokenizers library.
De Lauretis Tech
"""
import argparse
import logging
import os
import sys
from pathlib import Path
import yaml
from tokenizers import Tokenizer, models, trainers
from tokenizers.decoders import ByteLevel as ByteLevelDecoder
from tokenizers.normalizers import NFKC
from tokenizers.pre_tokenizers import ByteLevel
from tqdm import tqdm
from transformers import PreTrainedTokenizerFast
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(name)s - %(message)s",
datefmt="%Y-%m-%d %H:%M:%S",
)
logger = logging.getLogger("boccaccio.tokenizer")
BYTES_PER_GB = 1024 ** 3
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Train a BPE tokenizer for BoccaccioAI on CulturaX Italian data."
)
parser.add_argument(
"--config",
type=str,
default="configs/tokenizer.yaml",
help="Path to the tokenizer YAML configuration file.",
)
parser.add_argument(
"--output-dir",
type=str,
default=None,
help="Override the output directory from config.",
)
parser.add_argument(
"--corpus-size-gb",
type=float,
default=None,
help="Override the corpus size (in GB) used for training.",
)
parser.add_argument(
"--vocab-size",
type=int,
default=None,
help="Override the vocabulary size from config.",
)
return parser.parse_args()
def load_config(config_path: str) -> dict:
config_file = Path(config_path)
if not config_file.exists():
logger.error("Config file not found: %s", config_path)
sys.exit(1)
with open(config_file, "r", encoding="utf-8") as f:
config = yaml.safe_load(f)
logger.info("Loaded config from %s", config_path)
return config
def corpus_iterator(corpus_size_gb: float, source: str, lang: str):
"""Stream text samples from CulturaX until the target byte size is reached."""
from datasets import load_dataset
target_bytes = int(corpus_size_gb * BYTES_PER_GB)
logger.info(
"Streaming %s (lang=%s), target size: %.2f GB (%s bytes)",
source, lang, corpus_size_gb, f"{target_bytes:,}",
)
dataset = load_dataset(source, lang, split="train", streaming=True)
collected_bytes = 0
pbar = tqdm(
total=target_bytes,
unit="B",
unit_scale=True,
desc="Collecting corpus",
)
try:
for example in dataset:
text = example["text"]
text_bytes = len(text.encode("utf-8"))
collected_bytes += text_bytes
pbar.update(text_bytes)
yield text
if collected_bytes >= target_bytes:
break
finally:
pbar.close()
logger.info(
"Corpus collection complete: %.2f GB collected",
collected_bytes / BYTES_PER_GB,
)
def build_tokenizer(
vocab_size: int,
special_tokens: list[str],
min_frequency: int,
show_progress: bool,
) -> tuple[Tokenizer, trainers.BpeTrainer]:
"""Construct a byte-level BPE tokenizer and its trainer."""
tokenizer = Tokenizer(models.BPE())
tokenizer.normalizer = NFKC()
tokenizer.pre_tokenizer = ByteLevel()
tokenizer.decoder = ByteLevelDecoder()
trainer = trainers.BpeTrainer(
vocab_size=vocab_size,
min_frequency=min_frequency,
special_tokens=special_tokens,
show_progress=show_progress,
)
return tokenizer, trainer
def validate_tokenizer(tokenizer: Tokenizer, special_tokens: list[str]) -> None:
"""Print validation stats and fertility metrics for the trained tokenizer."""
vocab_size = tokenizer.get_vocab_size()
logger.info("Vocabulary size: %d", vocab_size)
for token in special_tokens:
token_id = tokenizer.token_to_id(token)
logger.info(" Special token: %-8s -> id %s", token, token_id)
test_sentences = [
"L'intelligenza artificiale sta trasformando il mondo della tecnologia.",
"Il gatto nero dorme tranquillamente sul divano di casa.",
"La Repubblica Italiana e' una democrazia parlamentare fondata sul lavoro.",
]
logger.info("--- Fertility analysis ---")
for sentence in test_sentences:
encoding = tokenizer.encode(sentence)
num_tokens = len(encoding.ids)
num_words = len(sentence.split())
fertility = num_tokens / num_words if num_words > 0 else 0.0
logger.info(" Text: %s", sentence)
logger.info(" Tokens: %d (words: %d, fertility: %.2f tok/word)", num_tokens, num_words, fertility)
logger.info(" Token ids: %s", encoding.ids)
logger.info(" Decoded: %s", tokenizer.decode(encoding.ids))
logger.info("")
def main() -> None:
args = parse_args()
config = load_config(args.config)
tok_cfg = config["tokenizer"]
# Resolve parameters (CLI overrides take precedence over config).
vocab_size = args.vocab_size or tok_cfg["vocab_size"]
corpus_size_gb = args.corpus_size_gb if args.corpus_size_gb is not None else tok_cfg["training_corpus_size_gb"]
output_dir = args.output_dir or tok_cfg["output_dir"]
output_name = tok_cfg["output_name"]
source = tok_cfg["training_corpus_source"]
lang = tok_cfg["training_corpus_lang"]
min_frequency = tok_cfg.get("min_frequency", 2)
show_progress = tok_cfg.get("show_progress", True)
special_tokens_cfg = tok_cfg["special_tokens"]
special_tokens = [
special_tokens_cfg["pad"],
special_tokens_cfg["bos"],
special_tokens_cfg["eos"],
special_tokens_cfg["unk"],
]
logger.info("=== BoccaccioAI Tokenizer Training ===")
logger.info("Vocab size: %d", vocab_size)
logger.info("Corpus size: %.2f GB", corpus_size_gb)
logger.info("Output dir: %s", output_dir)
logger.info("Special tkns: %s", special_tokens)
# Build tokenizer and trainer.
tokenizer, trainer = build_tokenizer(
vocab_size=vocab_size,
special_tokens=special_tokens,
min_frequency=min_frequency,
show_progress=show_progress,
)
# Stream corpus and train.
logger.info("Starting tokenizer training...")
text_iterator = corpus_iterator(corpus_size_gb, source, lang)
tokenizer.train_from_iterator(text_iterator, trainer=trainer)
logger.info("Training complete.")
# Save tokenizer.
output_path = Path(output_dir)
output_path.mkdir(parents=True, exist_ok=True)
json_path = output_path / f"{output_name}.json"
tokenizer.save(str(json_path))
logger.info("Saved raw tokenizer to %s", json_path)
# Save as HuggingFace-compatible PreTrainedTokenizerFast.
hf_tokenizer = PreTrainedTokenizerFast(
tokenizer_object=tokenizer,
bos_token=special_tokens_cfg["bos"],
eos_token=special_tokens_cfg["eos"],
unk_token=special_tokens_cfg["unk"],
pad_token=special_tokens_cfg["pad"],
)
hf_dir = output_path / f"{output_name}-hf"
hf_tokenizer.save_pretrained(str(hf_dir))
logger.info("Saved HuggingFace tokenizer to %s", hf_dir)
# Validate.
validate_tokenizer(tokenizer, special_tokens)
logger.info("Done.")
if __name__ == "__main__":
try:
main()
except KeyboardInterrupt:
logger.warning("Training interrupted by user.")
sys.exit(130)
|