tokenizer_module.py

"""HawkGPT 0.3 — Digit-aware BPE tokenizer.

Key improvement: numbers are split into individual digits.
"123" → ["1", "2", "3"] — model can learn arithmetic digit-by-digit.
"""

import os
from tokenizers import Tokenizer, models, pre_tokenizers, trainers

import config


def train_tokenizer(text_path: str, vocab_size: int = None) -> Tokenizer:
    """Train digit-aware BPE tokenizer."""
    if vocab_size is None:
        vocab_size = config.VOCAB_SIZE

    tokenizer = Tokenizer(models.BPE())

    # Digit-aware: Whitespace + individual digits
    tokenizer.pre_tokenizer = pre_tokenizers.Sequence([
        pre_tokenizers.Whitespace(),
        pre_tokenizers.Digits(individual_digits=True),
    ])

    trainer = trainers.BpeTrainer(
        vocab_size=vocab_size,
        special_tokens=["[PAD]", "[BOS]", "[EOS]", "[UNK]", "[MASK]"],
        min_frequency=3,
    )

    def line_iterator():
        with open(text_path, "r", encoding="utf-8") as f:
            for line in f:
                yield line

    tokenizer.train_from_iterator(line_iterator(), trainer=trainer)

    tokenizer.enable_padding(length=config.MAX_SEQ_LEN, pad_id=tokenizer.token_to_id("[PAD]"))
    tokenizer.enable_truncation(max_length=config.MAX_SEQ_LEN)

    os.makedirs(config.DATA_DIR, exist_ok=True)
    tokenizer.save(config.TOKENIZER_PATH)

    # Verify digit splitting
    tok_test = Tokenizer.from_file(config.TOKENIZER_PATH)
    tok_test.no_padding()
    tok_test.no_truncation()
    enc = tok_test.encode("123 + 456 = 579")
    print(f"Digit test: {'123 + 456 = 579'}")
    print(f"  Tokens: {enc.tokens}")
    print(f"  Digit-aware: {'1' in enc.tokens and '2' in enc.tokens and '3' in enc.tokens}")

    print(f"Tokenizer saved: {config.TOKENIZER_PATH} | vocab={tokenizer.get_vocab_size()}")
    return tokenizer


def load_tokenizer() -> Tokenizer:
    if not os.path.exists(config.TOKENIZER_PATH):
        raise FileNotFoundError(f"Tokenizer not found at {config.TOKENIZER_PATH}")
    return Tokenizer.from_file(config.TOKENIZER_PATH)


if __name__ == "__main__":
    tok = train_tokenizer(config.DATA_TEXT_PATH)
    tok.no_padding()
    tok.no_truncation()
    for test in ["123 + 456 = 579", "Привет! Как дела?", "Реши: 3x + 5 = 20"]:
        enc = tok.encode(test)
        print(f"  {test} → {enc.tokens}")