| """ |
| scripts/train_tokenizer.py |
| |
| CLI entry point for tokenizer training. Thin wrapper over |
| brain.tokenizer.tokenizer.BrainTokenizer.train() -- kept separate from the |
| tokenizer implementation itself so the implementation stays importable |
| without pulling in argparse/CLI concerns (same separation-of-concerns |
| principle as the model/API boundary elsewhere in this repo). |
| |
| Usage: |
| PYTHONPATH=. python3 scripts/train_tokenizer.py \\ |
| --corpus data/sample_corpus.txt \\ |
| --vocab-size 32768 \\ |
| --output data/brain_tokenizer.json |
| """ |
|
|
| import argparse |
|
|
| from brain.tokenizer.tokenizer import BrainTokenizer |
|
|
|
|
| def main() -> None: |
| parser = argparse.ArgumentParser(description="Train a Brain byte-level BPE tokenizer.") |
| parser.add_argument("--corpus", nargs="+", required=True, help="One or more corpus text files.") |
| parser.add_argument("--vocab-size", type=int, default=32768) |
| parser.add_argument("--min-frequency", type=int, default=2) |
| parser.add_argument("--output", required=True, help="Path to save the trained tokenizer JSON.") |
| args = parser.parse_args() |
|
|
| print(f"Training on: {args.corpus}") |
| print(f"Target vocab size: {args.vocab_size}") |
| tok = BrainTokenizer.train( |
| corpus_paths=args.corpus, |
| vocab_size=args.vocab_size, |
| min_frequency=args.min_frequency, |
| save_path=args.output, |
| ) |
| print(f"Saved to: {args.output}") |
| print(f"Actual vocab size: {tok.vocab_size}") |
| if tok.vocab_size < args.vocab_size: |
| print( |
| f"NOTE: actual vocab ({tok.vocab_size}) is smaller than requested " |
| f"({args.vocab_size}) -- the corpus doesn't contain enough distinct " |
| f"merges to reach the target. This is expected on small corpora; it " |
| f"is the same behavior the original Brain Trail tokenizer had " |
| f"(requested 500, got 344 on the 21-line sample corpus)." |
| ) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
